[T-Crypt]

What changed

Promotes only the post-v0.2 homelab results proven with repeats or quality-held-constant (main @ a02634f, v0.2 2026-10-08) into llm-tune's tables, findings, and traps. 7 files, +187/-3, SKILL.md steps still 1-9, no renumbering.

8 deltas promoted

1. Engine version is a tuning variable (alternating-arm A/B, quality held): Strata #783 +14.2% decode at 119K, +9.5% at 238K, recall 4/4; 0.1.43 agent-session +8.3%, prefill +6-8% at 16-32K; NInfer v3 prefill +1-2%, decode -0.8%, PPL +0.007% vs A16.
2. RAM-pinning trap numbers: 0.1.39 pinned 17.24->19.20 GiB, min free 3.36->1.41 GiB; headroom 6 restores 3.53 GiB (SKILL.md Step 4).
3. Short-prompt calibration vs depth (new finding 34): spec-min-p 0.70 costs decode at depth; short-prompt profile -20% at 119K; live is pcie-frac 0.20 + spec-min-p 0.5.
4. r14 head-to-head (finding 35): Strata ~35% faster decode, 2.1x faster 200K read, NInfer 1.7x richer output; review tie, recall 9/9 both. Report t/s AND quality.
5. Template/sampler token efficiency (n=1, preliminary): Sharp temp 0.6 same 10/12 in -18% tokens; presence 1.5 cuts MTP accept 89->84.2%, decode 308->254.
6. Chat-template effort trap (Step 7): v3 400s because jinja refuses effort high; fix maps high/max->xhigh, minimal->low.
7. Gemma4-26B-A4B-QAT DATA (2026-10-14, sm_89 native): 14.6 GiB weights, kv_headroom 0 @131072, decode 330-400 t/s, MTP 72-86%, tool-call ok. First non-Qwen family point.
8. Q4_K_XL fit-gate rejection: ruled out by fit gate before tuning ran (vram-fit T8).

Files (7)

• skill/references/tables/prefill-decode-speed.md (T9-T13)
• skill/references/tables/quality-bench.md (T16-T18)
• skill/references/tables/speculative-mtp.md (T7)
• skill/references/tables/vram-fit.md (T8)
• skill/references/findings.md (findings 33-35)
• skill/SKILL.md (traps only: Steps 4, 5/6, 7)
• data/INVENTORY.md (dated entries per source)

What the stages proved

• Tests: python -m py_compile skill/bench/*.py -> OK (all 3 scripts compile). pytest tests/ has no runnable tests (both subdirs are data: scenario answers, bench jsonl/txt) — nothing to break.
• Reviews clean: review.md found 2 minor nits (finding 34 untraced parenthetical; T18 missing denominator) + 2 cosmetic notes; review2.md verifies both nits FIXED (parenthetical removed, T18 header now raw-count non-comparable to T16). Table numbering sequential (prefill T1-T13, quality T1-T18, mtp T1-T7, vram T1-T8), all SKILL.md reference links resolve, CORRECTIONS.md untouched, engine tags on every new row, small-n items labeled preliminary.
• Arithmetic spot-checks (review stage, recomputed): T9 +14.3%/+9.5%, T10 +8.3%/+5.9%/+7.8%, T12 200K ~2.05x, v3 PPL +0.0067% — all match claims.
• Limitation (stated plainly): state/evals/ sources live in homelab, not this repo — figures checked for internal consistency + INVENTORY match, not against source logs.

Deliberately NOT promoted

• ninfer-pr15-norm/: kernel norm internals; no skill-level lesson identified.
• router-fast-tier/results.json, omp-pr2-ab.jsonl: no verdict read; needs a RESULTS.md first.
• r16-agentfast-parallel/: script only, never ran (no par.jsonl). Method reuse only.

Note on .router/

.router/2026-10-14-llm-tune-review/ (plan.md, review.md, review2.md) is untracked plan material and is NOT part of this PR. Repo convention treats planning docs as untracked (see commits ignoring docs/dev/), so it stays out unless the owner says to commit it.

Never merge without owner review — small-n items (r15 n=1, r14 n=2-3) are labeled preliminary and must not harden into rules without repeats.

[T-Crypt]

Independent review, 2026-10-14 (fresh pass; verdict: merge-with-nits)

Safety fact: the diff is exactly the 7 claimed .md files, zero non-md files, zero table-row deletions, SKILL.md steps still 1-9. Merging changes no executable behavior; the only risk channel is the numbers agents read, and the internal-consistency checks bound it.

Findings and their disposition:

1. FIXED this pass. T12 60K NInfer cell said "3,037 t/s (23.3 s)" but 55,586 / 3,037 = 18.3 s. Resolved against the run's own engine log (side-NInfer-Qwen3.8-27B.log, req#1): TTFT 18.3 s, total 23.3 s, prefill 3.04k tok/s. The logged seconds are whole-request wall time (answer generation included); the cell now says so and a note explains the semantics for the table.
2. FIXED this pass. README layout table said "lessons (32)"; findings.md now has 35. Updated to the counted value.
3. Noted, left as-is. T9 claims +14.2% at 119K; row values recompute to +14.33%. Within 0.1pp, consistent with 0.1 t/s input rounding; no rule derives from it.
4. Open nit. quality-bench T18 "gemma4-review (earlier pass)" row has no data/INVENTORY.md entry.
5. Open nit. INVENTORY.md section header lumps two dates.

Residual risk, stated plainly: source logs live in the operator's homelab, not this repo. Figures were verified for internal consistency and cross-checked against the engine log where it was in scope (T12), not re-derived from raw logs wholesale. Small-n rows (r15 n=1, r14 n=2-3) remain labeled preliminary and must not harden into rules without repeats.