feat(review): incremental re-review — review only new commits on synchronize, with a dedicated finding-resolution check [v0.9]

[devops-thiago]

Problem Statement

Every synchronize event re-reviews the full PR diff, no matter how small the push. Prompt size therefore grows with PR size, not push size, and the same already-reviewed hunks are re-sent (and re-billed) every round. Previous findings are condensed to fit (DiffBudgetPlanner.boundPreviousFindings, capped at 25% of the per-call budget — added after this repo's own release PR hit ~437K tokens in that block), but the diff itself has no such incremental treatment.

Proposed Solution

Bound follow-up rounds by push size:

• Persist the last-reviewed head SHA on ReviewSession. On synchronize, fetch GET /compare/{lastReviewedHead}...{newHead} instead of the full /pulls/{n}/files. Fall back to a full review whenever the compare fails or is misleading (force-push/rebase, base branch changed, no completed prior round) — the existing stale-head guard (ReviewContextLoader.java:498-515) already detects head movement mid-review.
• Dedicated resolution-check call, modeled on FindingVerificationService (blocking concise call, JSON verdicts keyed by id, fail-open, ledger-gated, salvage-on-truncation): input = condensed open findings (the entryLinesOnly shape from DiffBudgetPlanner.java:441: id, risk, file:line, title, plus suggestion_old) + only the new hunks touching those findings' files; output = resolved / unresolved / superseded per id. This takes over the previous_findings_status duty from the main review call, so the reviewer prompt no longer needs the reconciliation rule block (PrReviewPrompts.java:550-586) or the full previous-findings context — a second large prompt section shrinks.
• The review call runs over the incremental diff only → prompt size is bounded by the push, giving the predictable prompt sizes this roadmap is after. FollowUpAnalyzer thread matching, supersede logic, and the batch-scoping rules for statuses (FindingPipeline.scopeStatusesToBatch) are reused: the resolution-check call may only resolve findings whose touching hunks it actually saw.
• Escape hatches: manual /review keeps forcing a full-diff pass; a periodic full pass (e.g. every Nth round, or when ready_for_review fires) catches cross-commit interactions that incremental rounds can miss. Document the tradeoff.

Why a dedicated call beats the current reconciliation

Today reconciliation quality depends on the full re-review happening to look at the right places, and statuses must be carefully scoped per batch. A dedicated call sees exactly the pairs that matter (finding ↔ new hunks in its file) and nothing else — cheaper, more reliable, and it removes the largest conditional sections from the reviewer prompt.

Touchpoints

dashboard/ReviewSession (+ head SHA column/migration), review/ReviewContextLoader.java (compare-based incremental load + fallbacks), new review/ai/ResolutionChecker service + prompts (+ content test), review/FollowUpAnalyzer.java, review/FindingPipeline.java, review/DiffBudgetPlanner.java (call reservation), webhook/WebhookController.java (synchronize routing).

Acceptance criteria

• Push of K lines to an N-thousand-line PR produces prompts sized by K, not N (token metrics).
• Force-push/rebase and missing-history cases demonstrably fall back to full review.
• Resolution statuses match current behavior on the existing FollowUpAnalyzer test scenarios; a finding is never marked resolved by a call that didn't see its file's new hunks.
• Feature gate (REVIEW_INCREMENTAL_ENABLED), default off for one release.

Alternatives Considered

• Keep full-diff rounds and rely on tighter condensation — bounds the findings block but never the diff; cost still scales with PR size.