perf(review): scan backward for the last admissible label

.github/workflows/trusted-uv-materializer-quality-ci.yml

@@ -2,16 +2,15 @@ name: Trusted uv Materializer Quality CI
 
 on:
   pull_request:
+    branches: [main]
     paths:
       - ".github/workflows/trusted-uv-materializer-quality-ci.yml"
       - "scripts/ci/materialize_base_python_requirements.py"
-      - "scripts/ci/verify_release_maturin_tool_assets.py"
       - "tests/conftest.py"
       - "tests/test_materialize*.py"
       - "tests/test_trusted_uv*.py"
       - "tests/test_uv*.py"
       - "tests/test_repository_branch_coverage_*.py"
-      - "tests/test_verify_release_maturin_tool_assets.py"
       - "requirements-opencode-review-ci.txt"
       - "requirements-noema-document-ci.txt"
       - "requirements-opencode-review-ci-hashes.txt"
@@ -22,13 +21,11 @@ on:
     paths:
       - ".github/workflows/trusted-uv-materializer-quality-ci.yml"
       - "scripts/ci/materialize_base_python_requirements.py"
-      - "scripts/ci/verify_release_maturin_tool_assets.py"
       - "tests/conftest.py"
       - "tests/test_materialize*.py"
       - "tests/test_trusted_uv*.py"
       - "tests/test_uv*.py"
       - "tests/test_repository_branch_coverage_*.py"
-      - "tests/test_verify_release_maturin_tool_assets.py"
       - "requirements-opencode-review-ci.txt"
       - "requirements-noema-document-ci.txt"
       - "requirements-opencode-review-ci-hashes.txt"

CHANGELOG.md

@@ -1,16 +1,52 @@
-### Maturin download failures close every transport response
-
-- Refactor the bounded Maturin asset downloader so successful and rejected
-  responses share one unconditional close path while `HTTPError` keeps its own
-  explicit close path. A new regression exercises a non-200 response and an
-  opener-raised HTTP error. This removes an impossible optional-response branch
-  without changing hosts, redirects, byte limits, hashes, or fail-closed error
-  mapping; the focused suite is 17 passed with 100% statement and branch
-  coverage. The trusted full-suite workflow now also tracks the verifier source
-  and its focused test, so a future lifecycle change cannot omit the repository
-  coverage gate that detected this regression. The pull-request trigger admits
-  stacked canonical-owner bases as well as `main`; the protected-branch push
-  trigger remains restricted to `main`.
+### OpenCode invalid-token wrappers remain fail-closed
+
+- Treat a balanced object or array whose first token is invalid JSON as one
+  outer evidence container instead of skipping its opener and promoting a
+  nested exact-run control. RED commit
+  `b5a0507b8ee4ec70cc5fca751f34df98719f219b` binds `undefined`, arbitrary
+  unquoted array tokens, and unquoted object keys; GREEN commit
+  `443c29aa6f56735f7fcf0ee6a550b6bad5f7e2f8` preserves the proven prose
+  delimiter recovery while keeping balanced malformed wrappers fail-closed.
+  The production normalizer remains executable (`100755`).
+
+### OpenCode mixed-output framing preserves later controls
+
+- Ignore prose `{` or `[` delimiters whose next non-whitespace token cannot
+  start the corresponding JSON container. This preserves a later complete
+  exact-run control after diagnostics such as `Diagnostic: [pending` without
+  weakening the existing fail-closed rule for malformed outer objects or
+  arrays. RED commit `42099a359cdfb8126ea4e3d9118c60f6d98af1fa`
+  records the lost-control case; GREEN commit
+  `07baba20259c9061fbb647963f09745a9ca6931b` restores the framing boundary.
+  Fresh exact-head hosted Checks and a qualifying independent approval remain
+  required before ordinary merge.
+
+### OpenCode embedded JSON evidence is outermost and linear-time
+
+- Parse prose-wrapped JSON by scanning each outermost object or array once,
+  with string and escape awareness, instead of retrying `raw_decode` at every
+  nested `{`. A malformed outer container can no longer promote a nested
+  exact-run control object into top-level approval evidence, and adversarial
+  nested prefixes no longer cause quadratic decoding work. RED commit
+  `d05f67f6c5f164309d3cbf15f4858c5fa3d176cd` preserves the malformed-object
+  exploit; GREEN commit `86f52f320e8eb5ac56efdf9a3e2552b20bf409de`
+  replaces the retry loop, and `1a00c73a752b45d9d2a808d9bc96baae51ea8402`
+  extends the same invariant to malformed array nesting. Fresh exact-head
+  hosted Checks and a qualifying independent approval remain required before
+  ordinary merge.
+
+### OpenCode evidence labels require standalone identity boundaries
+
+- Reject a purported verification label when it is only the suffix of an
+  identifier-like token such as `uncoverage:`. The OpenCode normalizer now
+  applies the same boundary rule while selecting the last admissible label and
+  while finding the next label that terminates its section. RED commit
+  `af06cf9c87de4ac76db575d087a170746d6ab83d` proves that a forged suffix could
+  override an earlier fail-closed coverage statement; GREEN commit
+  `4d49b8307706ab8d4565cca8b0d9728bcdad2a35` preserves repeated labels,
+  Markdown decoration, and the distinct `docstring coverage:` label while
+  rejecting the identity-confused form. Fresh exact-head hosted Checks and a
+  qualifying independent approval remain required before ordinary merge.
 
 ### Shared Strix lock advances beyond the PyJWT recursion DoS
 

docs/product-technical-gap-baseline.md

@@ -1,11 +1,5 @@
 # Product and Technical Gap Baseline
 
-## 2026-10-01 Maturin response-lifecycle coverage closure
-
-| Gap | Exact evidence | Action | Status |
-|---|---|---|---|
-| `.github#1653@5cd141ec2c33b631d164af936cd1c9de70e4c9a4` passed all 5,314 tests but failed the complete branch gate because the canonical Maturin downloader left five error-path statements and two branches unexecuted; the owner workflow omitted both verifier paths and stacked PR bases | Trusted uv Materializer run `36811202519`, job `110206427182`; `verify_release_maturin_tool_assets.py` 95%, missing lines 104 and 106-112 plus branch 114→116; no owner run at #2530 predecessor `8cf2ea5f73976d47b2267fb52ac28284323404b7` because its base was #2531 rather than `main` | Repair canonical successor `.github#2530`: exercise non-200 and opener-raised `HTTPError` closure, replace the impossible nullable-response finalizer with one unconditional response-owned close scope, add source/test and stacked-PR trigger contracts to the complete gate while retaining protected-main push scope, preserve all network and fail-closed boundaries, then ordinary-merge the accepted owner head into #1653 | **Proposed / hosted RED reproduced; focused verifier coverage GREEN locally; path and stacked-admission contracts RED→GREEN; exact-head hosted full-suite, security, CodeQL, and independent approval required** |
-
 ## 2026-10-01 bounded Maturin release downloader SAST closure
 
 | Gap | Exact evidence | Action | Status |
@@ -3725,3 +3719,125 @@ verdict-shape acceptance, and qualifying independent approval.
 **Gap / failure scene.** The v2 handler names a run with `head/base/required-run/producer-source`, but its required-workflow fallback looked up only `head/base/required-run`. When authenticated status publication is unavailable, a completed clean handler job could not be found and a rerun ended false RED. Omitting the producer source would also allow a regenerated live merge revision to reuse predecessor evidence.
 
 **Action / evidence.** Correct the fallback lookup to include the live merge source and retain fail-closed base, head, required-run, workflow-path, job-name, GHAS-identity, and SARIF checks. The test-first repair reproduced two failures, then passed 96 focused workflow-contract tests; the new edge case rejects a stale merge-source title. Ruff E9/F/I on the changed dispatch-contract file and `git diff --check` pass. Fresh hosted Checks and a qualifying independent approval are still required on the unchanged executable delta before merge.
+
+## 2026-10-01 OpenCode evidence-label identity boundary
+
+**Status:** Proposed on `ContextualWisdomLab/.github#2543`; fresh exact-head
+hosted Checks and a qualifying independent approval remain mandatory.
+
+**Context Map / owner.** The central `.github` review-control bounded context
+owns OpenCode response normalization and approval-evidence admission. OpenCode
+is the untrusted evidence producer; repository review workflows consume only
+the normalizer's fail-closed verdict contract.
+
+**Gap / RCA.** The optimized last-label search used an unconstrained
+`rfind("coverage:")`. A response could state that real coverage evidence was
+not measured, then append `uncoverage: ... 100%`; the suffix beginning inside
+`uncoverage:` was accepted as a newer `coverage:` label. The false label could
+therefore replace the genuine fail-closed section and make
+`mentions_full_coverage` return true. The same missing identity boundary also
+affected forward searches for the next section label.
+
+**RED → GREEN / action.** RED
+`af06cf9c87de4ac76db575d087a170746d6ab83d` adds the durable suffix-forgery
+case. GREEN `4d49b8307706ab8d4565cca8b0d9728bcdad2a35` rejects label occurrences whose
+preceding character is alphanumeric, underscore, or hyphen in both backward
+selection and forward section termination. It retains decorated Markdown
+labels, repeated legitimate labels, and the separate `docstring coverage:`
+rule. Local direct behavior cases, Python compilation, and `git diff --check`
+are GREEN; the local environment has no pytest installation, so no full-suite
+claim is made. Completion still requires hosted exact-head tests, terminal
+required Checks, no unresolved actionable thread, qualifying independent
+approval, and ordinary protected integration.
+
+## 2026-10-01 OpenCode outermost JSON evidence boundary
+
+**Status:** Proposed on `ContextualWisdomLab/.github#2543`; executable repair
+and local focused verification are GREEN, while fresh exact-head hosted Checks
+and a qualifying independent approval remain mandatory.
+
+**Context Map / owner.** The central `.github` review-control bounded context
+owns OpenCode response normalization and exact-run evidence admission. The
+model response is untrusted input; downstream required-review workflows may
+consume only top-level controls accepted by this owner contract.
+
+**Gap / RCA.** Exact-head Strix run
+[36794394865](https://github.com/ContextualWisdomLab/.github/actions/runs/36794394865)
+reported repeated JSON decoding in `iter_json_objects`. Direct reproduction
+confirmed two effects from the same retry-at-every-`{` loop: a malformed outer
+object could promote its valid nested exact-run control into top-level evidence,
+and doubling an unclosed nested prefix increased processing time by roughly
+four times. The first effect is an identity-boundary false admission; the
+second permits model-controlled quadratic work.
+
+**RED → GREEN / action.** RED
+`d05f67f6c5f164309d3cbf15f4858c5fa3d176cd` records the malformed-object
+promotion. GREEN `86f52f320e8eb5ac56efdf9a3e2552b20bf409de`
+tracks one outermost container at a time, handles quoted strings and escapes,
+and decodes only a completed outermost span. Follow-up test commit
+`1a00c73a752b45d9d2a808d9bc96baae51ea8402` binds the same rule to malformed
+array nesting. The existing embedded-object cases and both new adversarial
+cases pass; Python compilation and `git diff --check` are GREEN. A direct
+3,200-level measurement completed in 0.001236 seconds, while the prior
+1,600-level case required 0.065367 seconds. These are local diagnostic values,
+not a hosted performance claim. Pytest is unavailable in the local runner, so
+the complete suite remains an exact-head hosted acceptance requirement.
+
+
+## 2026-10-01 OpenCode mixed-output JSON framing boundary
+
+**Status:** Proposed on `ContextualWisdomLab/.github#2543`; executable
+RED→GREEN complete, fresh exact-head hosted Checks and a qualifying independent
+approval remain mandatory.
+
+**Context Map / owner.** The central `.github` review-control bounded context
+owns mixed-output framing and exact-run evidence admission. OpenCode output is
+untrusted: prose delimiters must not block a later complete top-level control,
+while actual malformed outer containers must continue to suppress nested
+controls.
+
+**Gap / RCA.** CodeRabbit review on prior exact head
+`2ee8ddad072fe59bbe33c99994fa1e8c9cb5387c` showed that
+`Diagnostic: [pending` was treated as an unclosed JSON array and hid a later
+valid control. Direct reproduction showed the same regression for
+`Diagnostic: {pending`. The outermost single-pass repair had dropped the
+predecessor's candidate-start grammar gate.
+
+**RED → GREEN / action.** RED
+`42099a359cdfb8126ea4e3d9118c60f6d98af1fa` binds both prose delimiters.
+GREEN `07baba20259c9061fbb647963f09745a9ca6931b` admits an object start only
+before `"` or `}`, and an array start only before a JSON value starter.
+Existing malformed object/array nested-control cases remain fail-closed.
+Direct focused cases 8/8, Python compilation, and `git diff --check` are GREEN;
+pytest is unavailable locally, so the full hosted suite remains required.
+
+
+## 2026-10-01 balanced invalid-token wrapper evidence boundary
+
+**Status:** Proposed on `ContextualWisdomLab/.github#2543`; executable
+RED→GREEN is published, while fresh exact-head hosted Checks and a qualifying
+independent approval remain mandatory.
+
+**Context Map / owner.** The central `.github` review-control bounded context
+owns OpenCode mixed-output normalization and exact-run evidence admission.
+OpenCode model output is untrusted. A nested control inside any balanced outer
+container—including a syntactically invalid one—must never acquire top-level
+identity.
+
+**Gap / RCA.** The candidate-start allowlist skipped balanced wrappers whose
+first token was invalid JSON. Inputs such as `[undefined, {control}]`,
+`[unquoted_token, {control}]`, and `{unquoted_key: {control}}` therefore
+promoted the nested exact-run control. The earlier `NaN`/`Infinity` repair
+covered Python JSON extensions but not the general invalid-token boundary.
+
+**RED → GREEN / action.** RED
+`b5a0507b8ee4ec70cc5fca751f34df98719f219b` reproduces all three promotions.
+GREEN `443c29aa6f56735f7fcf0ee6a550b6bad5f7e2f8` treats an invalid starter as a
+malformed outer container when its token reaches a structural separator before
+another opener; the proven unclosed prose-delimiter recovery remains intact.
+Focused parser verification is 118/118 GREEN. Full repository verification
+reached 5,318 passed, 7 skipped, and 40 subtests; its sole failure was the
+published-commit ancestry test because the isolated `git archive` intentionally
+has no `.git` directory. Warnings-fatal compilation and diff whitespace checks
+are GREEN. Exact Git tree `dccb57a152c21600305c0eabac06cb29de2e0cf7`
+preserves source mode `100755`.