feat(review): persist every AI review for accuracy-over-time tracking

[JSONbored]

Priority: P1 (review-quality). Child of #1029.

Problem: AI reviews are not durably stored — ai_review_cache has 0 rows and stores only {notes, reviewer_count}. We cannot gauge review accuracy over time (verdict vs realized PR outcome).

Requirements: Persist the FULL review per (repo, pr, head_sha, mode): the verdict (consensusDefect/split/inconclusive/clean), blockers, nits, notes, readiness, disposition, model, timestamp. Wire the verdict↔pr_outcome correlation (computeGateEval / the accuracy loop) so a maintainer dashboard can show precision over time. Deliverables: a reviews/verdict store (migration), the write on every completed review (webhook + sweep paths), and the accuracy read. Relates to the #1462 cache verdict-drop (the cache should carry the verdict, not just notes).

[JSONbored]

Resolved by #1492 (fix(review): preserve cached AI gate findings), which added ai_review_cache.findings_json (migration 0078) and replays the gate findings on a cache hit + returns findings from runAiReviewForAdvisory. That fully covers the verdict-drop and persists the structured verdict per (repo,pr,head_sha). The running engine was rebuilt to current main so the column is now live. Closing as superseded.