Make code review incremental, risk-routed and measurable

Scope

Make code review converge with risk-based model routing, impact-based follow-up, consistent rules and measurable review cost. Implement all code-review recommendations approved by the maintainer as a separate PR. This is a process change, not an assertion that cheaper models have already passed evaluation.

Technical spec

• Initial audits use the actual PR base/merge-base, never origin/main for stacked changes. Fix audits and re-reviews use the last reviewed commit, expanding only for affected contracts and defect classes. Propagate explicit immutable base/head/previous SHAs through delivery and implementation inputs.
• Claims reviewers read manuals for context but block only changed claims and claims the change makes false; existing unrelated debt is separate. Preserve accepted future-state ADR documentation.
• Resolve precedence centrally: binding rules/accepted ADRs constrain specs; spec exceptions require explicit authorized rule/ADR amendments, not silent override. Required coverage includes always-in-scope/testing rules even if not repeated by spec. Budget exhaustion is incomplete assessment, neither defect nor pass; propagate it through reports and gate workflow.
• Replace src/text follow-up classification with impact: behavior, test logic, acceptance/spec, claims, and mechanical-only edits. Changed assertions get targeted recheck; claims runs only when affected; spec reruns when acceptance changes and has read-only source access.
• Keep one full review, targeted round 2, and final targeted fix check; unresolved important PR regressions at the cap park for escalation. Late pre-existing issues may be deferred, late PR regressions cannot silently be labelled unrelated.
• Deduplicate before verification using stable finding IDs, maintain resolution history and distinguish duplicate/incorrect/intentional/nonblocking outcomes. Verify each distinct defect once; consolidate fixes by defect class.
• Ordinary spec review Sonnet/medium, consequential ADR/cross-task decisions Opus/high. Claims Sonnet/medium. Coordination Sonnet/medium with Opus escalation for ambiguous consequential dismissals. Code remains Opus/high by default pending pilot; optional explicitly selected Sonnet/medium pilot for routine changes only. Ownership, destruction, persistence/recovery, concurrency, trust-boundary and uncertain risk require Opus regardless of diff size. Additional test lens is selected for test complexity/risk rather than only 1500 lines. No mandatory Haiku semantic reviewer.
• Add an executable/usable evaluation and measurement path: per-invocation role/model/commit/time/token categories/cost (unknown explicitly when unavailable), fix effort, all checks including fix-check, deduplicated finding outcomes, escaped defects. Preserve historical stats compatibility. Define representative buggy and clean PR replay, shadow pilot, severity-weighted misses and false positives, promotion criteria and rollback. Do not claim a completed empirical pilot.
• Update agent definitions, briefs, deliver/review and maintenance skills, delivery/testing-related rules, maintainer docs, scripts/hooks/report parsing and meaningful tests together. Keep implementation scope focused; no production changes. Exploration generic evidence contract is the other PR's responsibility.

Facts

• F1: delivery rule 14 already has parallel reviewers and at most two rounds plus fix check.
• F2: .agents/briefs/implement-audit.md uses git diff origin/main; implement-fix.md repeats that audit.
• F3: .claude/agents/reviewer.md classifies test-only fixes as text, bypassing re-review.
• F4: .claude/skills/deliver/SKILL.md excludes fix-check from round count and records whole-session usage.

Rules in play

• docs/internal/agent-rules/{delivery,testing,toolchain,documentation,always-in-scope}.md: amend consistently as authorized by this task; preserve required safety and behavior coverage.

Tests

Use existing script test surfaces (src/review-inputs.test.ts, src/delivery-stats.test.ts, src/pr-body-check.test.ts and related tests) and focused unit commands from toolchain.md. Prove stack bases and missing previous SHA fail closed where implemented; impact/incomplete report handling; every invocation counted; duplicate outcomes excluded from false-positive calibration; legacy stats remain readable. Validate process scenarios for complex tiny diffs, test-only fixes and late PR regressions.

Done when

• Incremental fix audits and actual PR bases are consistently propagated.
• Precedence, incomplete assessment, impacts and late-regression escalation are unambiguous across callers and reports.
• Risk/model routing and bounded targeted follow-up are usable; unproven code-model downgrade remains opt-in.
• Metrics and evaluation instructions support measuring quality, total elapsed time and total cost without inventing unavailable data.
• Required focused tests and direct spec/code/claims checks pass (maintainer explicitly requested no sub-agents); a separate PR is ready for maintainer review.

Out of scope

Production device behavior; blanket model downgrade without evaluation; merging this PR.

Depends on

None. Exploration changes are a separate independent PR.

Written by an agent.