review: Keep incomplete execution out of completed review caches

[idy]

Summary

A reviewer that could not read its supplied inputs returned policy blockers, and the workflow cached them as a completed review. New requests on the same head then reused the failure instead of retrying (observed in GizClaw/gizos#257).

Require an explicit execution status on every model turn, preflight supplied input paths, and instruct the reviewer to attempt read-only access even when the inputs live outside its working directory. Incomplete or invalid execution fails before the turn is checkpointed. Completed reviews with legitimate blockers remain reusable. Ledger schema v4 invalidates legacy evidence and chunks once; the published review payload and permission restrictions remain unchanged.

Fixes #35

Validation

• PASS: node .github/scripts/pr-review/test.mjs — failures and same-head recovery for PR, Issue, code chunk, and aggregation; missing status; real blocker reuse; legacy cache invalidation.
• PASS: The new incomplete-execution regression fails against the old runner (0 !== 1: it incorrectly returns success) and passes with the fix.
• PASS: node .github/scripts/pr-readiness/test.mjs
• PASS: node .github/scripts/issue-review/test.mjs
• PASS: node .github/scripts/review-request/test.mjs
• PASS: git diff --check

The tests exercise orchestration with a mocked Codex process. They do not establish live model obedience. Pinned consumers need to adopt the upstream revision after merge before the fix takes effect.

[idy]

@codex review

[github-actions]

❌ OpenAI PR Review: FAIL

Conclusion: Not ready. 2 readiness blockers and 1 actionable finding must be resolved before merge.

Review checks

Check	Result
PR format	✅ PASS
Issue design	✅ PASS
Code & plan conformance	❌ FAIL

Scope: aadfc5ba82 · cb2d133738..aadfc5ba82 · full · 1 diff chunk

Usage: 1m 10s · 301,036 tokens · 77.9% cache hit · 7.071 credits

Blockers

• plan-conformance Issue #35 requires incomplete execution to report the concrete inability. The implementation accepts an empty incomplete reason, leaving failed runs without that required diagnostic.
• code-review .github/scripts/pr-review/run.mjs:207: Reject incomplete responses without a reason

Findings

1 inline finding published.

Summary

The incomplete-execution path accepts an empty reason, so an execution failure can be reported without the concrete cause required by the Issue.

Review metadata

• Commit: aadfc5ba82
• Range: cb2d133738..aadfc5ba82
• Mode: full
• Diff chunks: 1
• Model: gpt-5.6-terra
• Reasoning effort: medium
• Session: repo-1309321116-pr-36-v2
• Generation: d84756da1e2d884878ae41d184772f2dbe8207506c2db4dd0f171f33df389ba1
• Evidence: f86e23ea9de4dc587c8a21a7541986e23cacba182dd7f171b3e29389d8ea271c

Totals

• Input: 296,962
• Cached input: 231,415
• Cache write: 65,508
• Output: 4,074
• Reasoning: 1,660
• Total: 301,036
• Estimated credits: 7.071
• Credit rate per 1M tokens: 62.5 input / 6.25 cached / 375 output

Token and cache details

Stage	Mode	Target	Time	Input	Cached	Hit	Output	Total	Credits
pr	deterministic	format rules	0s	0	0	0.0%	0	0	0.000
pr	full	pr	14s	29,165	18,843	64.6%	481	29,646	0.943
issue	full	Issue #35	17s	62,340	50,566	81.1%	932	63,272	1.401
code	full	chunk 1/1	33s	120,559	80,755	67.0%	2,249	122,808	3.836
code	full	aggregate	6s	84,898	81,251	95.7%	412	85,310	0.890

[github-actions]

✅ OpenAI PR Review: PASS

Conclusion: Ready from the OpenAI review perspective. PR format, linked Issue design, and code/plan conformance passed with no actionable findings.

Review checks

Check	Result
PR format	✅ PASS
Issue design	✅ PASS
Code & plan conformance	✅ PASS

Scope: f102acbd8e · aadfc5ba82..f102acbd8e · incremental · 1 diff chunk

Usage: 15s · 186,037 tokens · 97.7% cache hit · 1.700 credits

Summary

The prior incomplete-reason finding is demonstrably fixed by the incremental change and its regression tests. No actionable findings remain.

Review metadata

• Commit: f102acbd8e
• Range: aadfc5ba82..f102acbd8e
• Mode: incremental
• Diff chunks: 1
• Model: gpt-5.6-terra
• Reasoning effort: medium
• Session: repo-1309321116-pr-36-v2
• Generation: 35e763ed12522777e678148363c65df14fb4290c8c398ae1f11230a42c1006fa
• Evidence: 22adc5135a8dc56072d7a91d98962411b5752b73823f887d2104d9e0a9d14e98

Totals

• Input: 185,238
• Cached input: 180,926
• Cache write: 4,300
• Output: 799
• Reasoning: 192
• Total: 186,037
• Estimated credits: 1.700
• Credit rate per 1M tokens: 62.5 input / 6.25 cached / 375 output

Token and cache details

Stage	Mode	Target	Time	Input	Cached	Hit	Output	Total	Credits
pr	deterministic	format rules	0s	0	0	0.0%	0	0	0.000
pr	reused	cached evidence	0s	0	0	0.0%	0	0	0.000
issue	reused	Issue #35	0s	0	0	0.0%	0	0	0.000
code	incremental	chunk 1/1	10s	90,711	87,757	96.7%	580	91,291	0.951
code	incremental	aggregate	5s	94,527	93,169	98.6%	219	94,746	0.749