ai(cache): review fingerprint omits prompt and verdict-logic version — superseded verdicts replay forever (11 unbumped changes)

[JSONbored]

Summary

getCachedAiReview replays a cacheable=1 row unboundedly (there is no maxAgeMs), and the cache fingerprint does not include the reviewer prompt, the verdict-combination logic version, or several per-repo flags that change what the model is asked. So a verdict produced under superseded policy is replayed verbatim — including a severity: "critical" ai_consensus_defect — and the fixed code never runs for that head.

Mechanism (verified at HEAD, 776c414)

• src/db/repositories.ts:5331-5366, :5385-5415 — getCachedAiReview and getCachedAiReviewAcrossHeads reuse durable rows with no age bound; stored findings are replayed as-is.
• AI_REVIEW_CACHE_INPUT_VERSION is v6 (src/review/ai-review-cache-input.ts:17), last bumped for AI CLI timeout knobs are env-var-only, unlike the sibling model/effort knobs (config-as-code gap) #8364. Since then fix(ai-review): the incoherent-diff bail misfires on title-vs-diff scope mismatch, turning reviewable PRs into inconclusive holds #8789, fix(ai-review): inconclusive-finding copy claims 'dual-model' in single+fallback mode, and claude_code auth errors lack structural circuit-breaker classification #8791, gate: move every objectively-checkable criterion from LLM judgment to deterministic checks #8833, fix(orb): enforce the model/fact boundary — CI claims demoted at parse, silence is not certainty #8845, Review-input truncation causes false 'missing screenshot evidence' closes #8961, orb(security): AI reviewer prompt has no instruction-hierarchy defense — untrusted PR text is concatenated raw #9035, orb(ai-review): 'AI reviewers agree' is never verified — consensus fires whenever BOTH have any blocker, not the same one (427 closes) #9074, orb(ai-review): a model that declared it could NOT review the diff can still emit blockers that close the PR #9087, fix(ai-review): verify reviewers actually agree; strip blocker authority from a bailed review #9114, fix(ai-review): fence untrusted PR content, stop shifting inline anchors, and stop judging a fix you cannot see #9145, fix(review): filter the AI narrative assessment per sentence, not all-or-nothing #9445 all changed prompt text or verdict logic without a bump.
• REVIEW_PROMPT_VERSION (src/services/ai-review.ts:1290) is still "review-prompt-v1" despite its own doc instructing a bump on any verdict-surface change — which additionally means counterfactual: advisory check when reviewer prompt versions change #8222's counterfactual replay gate silently no-ops.
• systemPromptDigest is computed (:2720) but reaches only the decision record, never the cache key.

Also absent from the fingerprint: findingCategories and improvementSignal (both append system-prompt suffixes — :1268, :1278), screenshotEvidenceSummary (appended to the system prompt, :1320), AI_MAX_OUTPUT_TOKENS, and the operator plan's onMerge (reviewerPlan hashes only combine and reviewers).

Concrete failure

A PR is closed or held under the pre-#9074 "false consensus" logic. #9074 lands, requiring genuine agreement. The PR is re-gated at the same head — identical fingerprint, cache hit — so the old critical finding and its close rationale are replayed and the corrected logic never executes. The fix is invisible for every already-reviewed head.

Requirements

1. Anything that changes what the model is asked, or how its answer is judged, must change the cache key.
2. Correctness must not depend on a human remembering to bump a constant — that mechanism has already failed 11 times.

Deliverables

• Fold REVIEW_PROMPT_VERSION and sha256(buildCanonicalJudgePrompt()) into the cache payload. Both are pure and available before the call, so prompt drift becomes an automatic miss rather than a manual bump.
• Add the missing per-repo/plan inputs: findingCategories, improvementSignal, screenshotEvidenceSummary, AI_MAX_OUTPUT_TOKENS, onMerge.
• Bump AI_REVIEW_CACHE_INPUT_VERSION once as part of this change, to invalidate rows written under the incomplete fingerprint.
• Consider a maxAgeMs backstop on durable reuse, so a fingerprint gap can never mean "forever".
• Add a drift check: fail CI when a file that contributes to the prompt changes without the digest mechanism covering it — or, better, make the digest structural so the check is unnecessary.

Tests (must fail against current main)

• Same head, prompt text changed ⇒ cache miss.
• Same head, findingCategories / improvementSignal / onMerge toggled ⇒ cache miss for each.
• Same head, nothing changed ⇒ cache hit (no regression in hit rate for the genuine case).
• A row written under v6 is not served after the bump.

Expected outcome

A policy or prompt fix takes effect on the next pass for every PR, instead of being masked by a cache row that predates it.

[linear-code]

JSO-6541