fix(loop-memory): canonicalize trusted check reruns

.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/CHUNK_MAP.md

@@ -2,10 +2,12 @@
 
 | Order | Chunk | Purpose | Risk | State |
 |---:|---|---|---:|---|
-| 0 | `WS-ENG-007-00R1` | Repair planning-intake file/tree parity and recover PR #187 exactly once | L1/P0 | Emergency prerequisite review |
-| 1 | `WS-ENG-007-01` | Add deterministic reviewed-patch identity and conservative base-delta review preservation | L1 | Blocked on 00R1 merge, signed reconciliation, and explicit start |
+| 0 | `WS-ENG-007-00R1` | Repair planning-intake file/tree parity and recover PR #187 exactly once | L1/P0 | Merged; recovery superseded after rerun-cardinality failure |
+| 0 | `WS-ENG-007-00R2` | Canonicalize repeated trusted check evidence and reconcile PRs #187 and #188 exactly once | L1/P0 | Proposed reliability recovery |
+| 1 | `WS-ENG-007-01` | Add deterministic reviewed-patch identity and conservative base-delta review preservation | L1 | Blocked on 00R2 merge, successful three-merge reconciliation, and explicit start |
 | 2 | `WS-ENG-007-02` | Add structured reviewer-track and upstream-finding reconciliation | L1 | Blocked on 01 merge and explicit start |
 | 3 | `WS-ENG-007-03` | Add merge-group CI parity and queue-readiness proof | L1 | Blocked on 02 merge and explicit start |
 
-Each chunk is one PR and stops after merge. Every successor requires a separate
+Recovery chunks are exceptional ordered prerequisites; implementation chunks
+remain one PR each and stop after merge. Every successor requires a separate
 explicit signed start.

.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/DECISIONS.md

@@ -62,3 +62,23 @@ Use the existing closed schema-v1 two-merge recovery mechanism, rebound to PR
 `WS-ENG-007-00R1`. Both ephemeral
 exemptions must be consumed before signing; no wildcard, manual state edit,
 force push, new secret, or persistent exemption is allowed.
+
+## D12 - Check reruns are evidence versions, not ambiguity
+
+For each protected check name, validate every returned same-name check against
+the exact reviewed head and pinned GitHub Actions application. Reject the set if
+any candidate is foreign, wrong-head, malformed, incomplete, or reuses an ID.
+Order completed candidates by parsed `started_at` instant then positive numeric
+check-run ID. The unique latest invocation determines the result. `completed_at`
+must be a valid instant at or after start, but completion order never defines
+recency. A later-started failure supersedes an older success even when the older
+run completes last. Rerun count alone never fails reconciliation.
+
+## D13 - Recover the exact three-merge backlog once
+
+Schema v3 names an ordered `recovered_merges` list of one or two entries plus
+one activation; production requires exactly two recovered entries.
+For this repair it binds only PR #187 / `WS-ENG-007-PLAN`, PR #188 /
+`WS-ENG-007-00R1`, and direct-next `WS-ENG-007-00R2`. The plan must be exactly
+those adjacent first-parent merges. All exemptions are consumed before signing
+and cannot appear in state, ledger, projections, or replay.

.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/DISCOVERY.md

@@ -59,3 +59,20 @@ tracks. A PR never authors or narrows this graph.
 - Exact GitHub merge-queue repository-setting API behavior.
 - Hosted context behavior after queue enablement. These block administrative
   enablement, not repository-side parity planning.
+
+## Recovery reliability discovery — 2026-07-23
+
+- PR #188 merged at `c65633f8f0991dbefe7b0635e053aab0df8f9af8`.
+- Loop Memory run `29984940789` failed before publication with `planning intake
+  check agent-gates is missing or duplicated`.
+- PR #187 head `34ddac158d8c8d5c96498de008dd43354205199f`
+  has two completed successful `agent-gates` runs from GitHub Actions app
+  `15368/github-actions` and one completed successful `test` run from the same
+  app.
+- `_validate_protected_actions_checks()` requires exactly one match, while
+  `_check_evidence()` already treats check history as a latest-value stream.
+- The current schema-v1 recovery certificate can name only one recovered merge
+  plus activation. Signed state is two merges behind, so the repair target is a
+  third adjacent merge and requires an exact ordered multi-recovered schema.
+- The workflow already plans all unrecorded first-parent commits and publishes
+  only after reconciliation; no workflow edit is required.

.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/INTENT.md

@@ -79,3 +79,34 @@ forged manifests, and merge-group combined trees.
 Approve each implementation chunk and the later repository-setting change that
 enables GitHub merge queue. No queue or automatic merge setting changes in the
 planning intake.
+
+## Recovery reliability addendum — WS-ENG-007-00R2
+
+### Problem being solved
+
+Signed reconciliation treats legitimate same-name GitHub Actions reruns as
+ambiguity. PR #187 therefore cannot enter signed history even though both
+`agent-gates` runs and its `test` run are trusted and successful.
+
+### Target behavior
+
+Rerun count is harmless. A deterministic latest trusted run controls each
+protected result, later failures cannot be hidden by older successes, and the
+exact three-merge backlog is reconciled atomically without reusable authority.
+
+### Boundaries preserved
+
+Required checks, branch protection, signing keys, permissions, human merge
+approval, coverage floors, product behavior, and successor start gates do not
+change.
+
+### Proof strategy
+
+Adversarial check histories, permutation invariance, exact production policy
+pinning, ordered recovery consumption, replay rejection, and byte-identical
+idempotency evidence.
+
+### Human decision
+
+The user's explicit `start` instruction approves planning and execution of this
+bounded reliability repair only. Successor starts remain separately gated.

.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/PLAN.md

@@ -181,3 +181,28 @@ always:
 
 Stop if preservation depends on mutable PR prose, inaccessible commits, an
 unbounded semantic judgment, weakened CI, or automatic human approval.
+
+## Recovery reliability plan — WS-ENG-007-00R2
+
+1. Replace exact-cardinality protected-check validation with a deterministic
+   latest-run selector that validates all same-name candidates against exact
+   head, pinned app, structural fields, and parseable timestamps before sorting.
+2. Reject the complete name-set if any candidate is malformed, untrusted, or
+   incomplete. Order completed candidates by parsed `started_at` instant then
+   positive numeric check-run ID. `completed_at` proves terminal consistency
+   but never defines invocation recency. Require the selected run to succeed.
+3. Extend the closed recovery certificate with schema v3: an ordered
+   `recovered_merges` list plus one activation. Validate exact plan equality,
+   chronological first-parent adjacency, identities, aggregate required checks,
+   protected-check provenance on every recovered/target head, uniqueness, and
+   consumption. Schema v3 permits at most two recovered merges; production
+   requires exactly both named entries.
+4. Bind production recovery to PR #187, PR #188, and direct-next 00R2 only.
+5. Prove adversarial rerun histories, permutation invariance, exact recovery,
+   non-serialization, replay rejection, and deterministic repeated output.
+6. Preserve the existing workflow, signing, human approval, successor stops,
+   and required-check names unchanged.
+
+Rejected alternatives: ignore duplicate runs, accept any successful run, use API
+array order, use only a check name/status context, rerun until history changes,
+manually edit signed state, or add a persistent recovery bypass.

.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/RISKS.md

@@ -12,3 +12,6 @@
 | Concurrent queue entries interact | High | Validate each exact merge-group SHA and invalidate on queue recomputation. |
 | Planning intake cannot enter signed history | Critical | Filter only validated directory entries from the GitHub recursive tree map, preserve every validated supported non-tree leaf identity, reject unsupported entries, and recover exact PR `#187` plus 00R1 through a consumed two-merge certificate. |
 | Recovery becomes a reusable bypass | Critical | Bind schema-v1 recovery to PR #187 merge `8928ba80eeaf31e609dbdeda7d2cc22e9ea482c8` and the exact 00R1 activation identity; require the ordered two-merge plan, successful checks, full consumption, and no persisted exemption. |
+| Legitimate check reruns block signed memory | Critical | Validate every same-name candidate, order completed trusted invocations by parsed start instant and positive check-run ID, then require the unique latest candidate to succeed. |
+| A stale success hides a later failure | Critical | Latest validated run wins; an in-progress, failed, cancelled, foreign-app, malformed, or timestamp-ambiguous latest run fails closed. |
+| Three-merge repair becomes general authority | Critical | Schema v3 permits at most two recovered merges and production names only PR #187, PR #188, and 00R2; require exact first-parent adjacency, unique identities, full consumption before signing, and no serialization. |

.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/STATUS.md

@@ -1,14 +1,16 @@
 # STATUS: WS-ENG-007 - Concurrent PR Review Reconciliation
 
-- Phase: planning-intake recovery
+- Phase: canonical check-evidence recovery
 - Gate: fail-closed automation repair
 - Active planning chunk: none
 - Active implementation chunk: none
-- Recovery chunk: `WS-ENG-007-00R1`
+- Completed but unreconciled recovery chunk: `WS-ENG-007-00R1`
+- Proposed recovery chunk: `WS-ENG-007-00R2`
 - Proposed implementation successor after recovery: `WS-ENG-007-01`
 - Separate explicit start required: true
-- Current gate: PR #187 merged at
-  `8928ba80eeaf31e609dbdeda7d2cc22e9ea482c8`, but signed post-merge memory and
-  the fresh explicit start both fail closed because recursive GitHub tree
-  evidence includes directory entries while the reviewed PR file inventory and
-  independent local Git checker contain only recursive non-tree entries.
+- Current gate: PR #188 merged at
+  `c65633f8f0991dbefe7b0635e053aab0df8f9af8`. Its tree normalization repair is
+  correct, but signed reconciliation still fails because PR #187 has two
+  successful trusted `agent-gates` reruns and planning intake incorrectly
+  treats repeated check evidence as ambiguous. Signed state remains at
+  `73b457925b02301587b83d01ced0adb66319d134`; no successor is active.

.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/chunks/WS-ENG-007-00R2-canonical-check-recovery.md

@@ -0,0 +1,148 @@
+# Chunk Contract: WS-ENG-007-00R2 — Canonical Check Evidence Recovery
+
+## Goal
+
+Make signed merge reconciliation stable under legitimate GitHub Actions reruns
+and reconcile the exact PR #187 → PR #188 → 00R2 backlog once.
+
+## Why this chunk exists
+
+PR #188 fixed recursive tree parity, but run `29984940789` failed because PR
+#187 has two successful trusted `agent-gates` check runs. Cardinality is mutable
+GitHub history and is not a valid ambiguity signal. Signed state remains at
+`73b457925b02301587b83d01ced0adb66319d134`.
+
+## Start phase
+
+`implementation`
+
+## Risk
+
+L1 / P0 policy, audit, CI-evidence, and signed-memory recovery
+
+## Authorization boundary
+
+This otherwise-unstartable repair uses a reviewed exact schema-v3 recovery
+certificate. It grants no workflow, repository, secret, merge, or continuing
+start authority.
+
+## Allowed files
+
+```text
+scripts/update_post_merge_memory.py
+scripts/test_update_post_merge_memory.py
+scripts/test_agent_gates.py
+scripts/test_check_loop_memory_state.py
+.agent-loop/policies/loop-memory-recovery.json
+docs/operations_post_merge_memory.md
+.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/CHUNK_MAP.md
+.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/STATUS.md
+.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/RISKS.md
+.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/DECISIONS.md
+.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/chunks/WS-ENG-007-00R2-canonical-check-recovery.md
+.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/reviews/WS-ENG-007-00R2-internal-review-evidence.md
+.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/reviews/WS-ENG-007-00R2-pr-trust-bundle.md
+.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/reviews/WS-ENG-007-00R2-external-review-response.md
+.agent-loop/merge-intents/WS-ENG-007-00R2.json
+```
+
+## Not allowed
+
+- Workflow, required-check name, branch-protection, coverage, dependency,
+  signing-key, secret, permission, product, backend, or frontend changes.
+- Treating CodeRabbit status as implementation-start authority.
+- Accepting a foreign app, wrong head, incomplete or malformed check run.
+- Letting an older success override a later failure.
+- Persistent, wildcard, reorderable, cross-repository, or reusable recovery.
+- Starting `WS-ENG-007-01` or `WS-ENG-006-01` in this chunk.
+
+## Acceptance criteria
+
+- [ ] Every same-name candidate is validated before selection. Any poisoned
+      candidate fails whether older or newer than a valid success and under
+      every input permutation. Completed trusted candidates are ordered by
+      parsed `started_at` instant then positive check-run ID; the unique latest
+      invocation determines success. Completion order never defines recency.
+- [ ] Zero matches, incomplete API response, foreign app, wrong head, missing or
+      malformed ID/timestamps, any non-completed run, non-success latest
+      conclusion, or a canonical identity collision fails closed. Timestamp
+      parsing is timezone-aware RFC3339 normalized to instants; naive/non-string
+      values fail. IDs require `type(id) is int` and `id > 0`; booleans,
+      byte-identical duplicate IDs, and conflicting ID reuse fail.
+- [ ] Tests cover one success, two successes, failure→success, success→failure,
+      an older-started success completing after a newer-started failure,
+      queued/in-progress evidence, cancelled/timed_out/skipped conclusions,
+      foreign app with the protected name, wrong head, offset-equivalent and
+      identical timestamps resolved by ID, duplicate ID rejection, malformed
+      timestamps, API truncation, and evidence order permutation. Poisoned
+      candidates appear both older and newer than a valid success. Later failure
+      is tested independently for `agent-gates` and `test`.
+- [ ] The real PR #187 shape—two successful trusted `agent-gates` plus one
+      successful trusted `test`—passes deterministically and is permutation
+      invariant.
+- [ ] Schema v3 permits only `schema_version`, ordered `recovered_merges`, and
+      `activation`; the list contains one or two entries, is identity-unique and SHA-unique,
+      and every entry binds initiative, chunk, positive PR number, and merge SHA.
+      Tests reject unknown/missing keys at every level, empty/three-entry lists,
+      duplicate identity or SHA under changed counterparts, bool/invalid PR,
+      malformed SHA, activation collision, and policy/Git order mismatch.
+- [ ] Production policy binds exactly PR #187 merge
+      `8928ba80eeaf31e609dbdeda7d2cc22e9ea482c8` / `WS-ENG-007-PLAN`, then PR #188
+      merge `c65633f8f0991dbefe7b0635e053aab0df8f9af8` /
+      `WS-ENG-007-00R1`, then activation `WS-ENG-007-00R2`.
+- [ ] Recovery requires the exact adjacent first-parent plan
+      `[8928ba80…, c65633f8…, target]`, validates every recovered identity and
+      required-check aggregate, validates exact protected-check provenance on
+      every recovered and target reviewed head, consumes all three exemptions
+      before signing, serializes none,
+      and rejects missing, extra, reordered, repeated, intervening, wrong-parent,
+      wrong-identity, check-failure, partial-consumption, and replay cases.
+- [ ] Ephemeral recovery-file schema v1 retains its original maximum of two
+      entries. Transport schema v2 is emitted only for the exact three-entry
+      schema-v3 recovery result, requires exactly three unique entries, and
+      rejects unknown versions, partial inventories, and four entries.
+- [ ] Applying the exact authenticated plan to two fresh copies of the same
+      starting state produces byte-identical complete generated closed trees,
+      including signing input. Replay from completed state proves no exemption
+      reinjection and no publication change.
+- [ ] Final state records 00R2 stopped with `WS-ENG-007-01` as explicit-start
+      successor; both `WS-ENG-007-01` and `WS-ENG-006-01` remain inactive.
+- [ ] Existing tree, planning grammar, merge attribution, lifecycle, signing,
+      checker, and recovery regression suites remain blocking and unchanged in
+      authority semantics.
+- [ ] Exactly one schema-v2 merge intent names `WS-ENG-007-01` as the same-
+      initiative explicit-start successor.
+
+## Verification commands
+
+```bash
+PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 python3 -m pytest -q scripts/test_update_post_merge_memory.py scripts/test_agent_gates.py scripts/test_check_loop_memory_state.py
+python3 scripts/test_agent_gates.py
+python3 scripts/update_post_merge_memory.py validate-merge-intent --repository-root . --base-ref origin/main
+PR_HEAD_SHA="$(git rev-parse HEAD)" python3 scripts/check_internal_review_evidence.py
+python3 scripts/check_markdown_links.py
+python3 scripts/check_stale_workstream_wording.py
+git diff --check origin/main...HEAD
+```
+
+## Required reviewers
+
+- senior engineering
+- QA/test
+- security/auth
+- product/ops
+- architecture
+- CI integrity
+- docs
+- reuse/dedup
+- test delta
+
+## Human review focus
+
+Confirm latest-run semantics match GitHub rerun behavior, later failures cannot
+be hidden, and schema-v3 authority is exact, ordered, consumed, and non-reusable.
+
+## Stop condition
+
+Stop after this PR. Do not start either successor until the user merges this
+specific PR and signed automation proves the exact recovered state.

.agent-loop/initiatives/WS-ENG-007-concurrent-pr-review-reconciliation/reviews/WS-ENG-007-00R2-external-review-response.md

@@ -0,0 +1,27 @@
+# External Review Response: WS-ENG-007-00R2
+
+## Comments addressed
+
+- CodeRabbit minor: reflowed the PR `#187` reference in the operations runbook
+  so Markdown does not parse it as a malformed ATX heading. The review thread
+  is resolved on PR #189.
+
+## Comments deferred
+
+None.
+
+## Human decisions needed
+
+None.
+
+## Commands rerun
+
+```bash
+python3 scripts/check_markdown_links.py
+python3 scripts/check_stale_workstream_wording.py
+git diff --check origin/main...HEAD
+```
+
+## Remaining risks
+
+No new risk. Exact adjacency and fresh exact-head checks remain blocking.