fix(review): gate the auto-tune breaker on reversal-weighted precision (#2348)

[JSONbored]

Summary

• planAutoTune/shouldAutoClear and their close-side mirrors (planCloseAutoTune/shouldAutoClearClose) now gate on GateEvalRow's reversal-discounted weightedMergePrecision/weightedClosePrecision (already computed by parity.ts's computeGateEval, landed in feat(review): discount reverted/reopened outcomes in the gate-eval fold #4169) instead of the raw precision fields — a high volume of later-reverted merges can no longer keep the raw number artificially healthy to dodge the breaker.
• The raw fields are preserved on AutoTuneAction/CloseAutoTuneAction for log continuity, and both numbers are surfaced in the alert message (weighted merge precision 50% (raw 95%) ...) so an operator can see the gap at a glance.
• computeTuningRecommendations (the separate advisory-only tuning surface, not "the breaker") is deliberately unchanged — out of this issue's named scope.
• evalRowFromCalibration (selftune-wire.ts, sourced from agent_recommendation_outcomes) has no reversal signal in its data source, so its weighted fields mirror the raw ones by construction — documented inline.

Closes #2348.

Re-verification of the issue's second downstream consumer (orb collector)

The issue's deliverable 4 names two downstream consumers to re-verify: the auto-tune breaker (this PR) and the orb collector (src/orb/analytics.ts). I checked — the orb collector's foldInstance does not consume parity.ts's GateEvalRow/computeGateEval at all; it folds its own independent mergePrecision directly from fleet-submitted orb_signals cells, and already excludes a reversal_flag !== "reverted" merge from mergeConfirmed (see the field's own doc comment: P(merged & not reverted | gate said merge)). It predates and is structurally separate from this weighted-fold work, and was already reversal-aware by construction. No change needed there.

Scope

• The PR title follows type(scope): short summary Conventional Commit format.
• This PR is focused (2 src/ files, 3 test files) and does not mix unrelated changes.
• Follows CONTRIBUTING.md; no site//CNAME/VitePress touched.
• Linked open issue: Closes maintainer: value-weighted gate-prediction calibration — optimize for merged-net-positive, not raw volume #2348.

Validation

• npm run typecheck — clean.
• npm run test:coverage (full unsharded) — src/review/auto-tune.ts 100% lines (83/83) / 100% branches (79/79). src/review/selftune-wire.ts 100% lines (30/30); its one uncovered branch (line 123, isAgentConfigured guard in the pre-existing, untouched selfTuneRepos) is outside this diff — confirmed via raw lcov.info BRDA inspection that both new lines (including the new wouldMerge > 0 ? ... : null ternary) are fully hit.
• Targeted suites: auto-tune.test.ts + outcomes-wire.test.ts + selftune-wiring.test.ts + selftune-readback.test.ts — 136/136 passing, including new tests proving the actual anti-gaming property (raw precision healthy + weighted precision failing → breaker now engages where it previously wouldn't; and the inverse recovery case for shouldAutoClear/shouldAutoClearClose).
• npm run test:workers / npm run build:mcp / npm run ui:* — not run locally; no MCP, worker-binding, wrangler, or UI surface touched by this diff (pure src/review/** logic + tests), left to CI per this session's established "typecheck + affected tests locally, trust CI for the rest" practice.
• npm audit --audit-level=moderate — no dependency changes in this PR.
• New branches (the ?? raw-precision fallback on both the merge and close side) and the core anti-gaming property have dedicated tests, not just line coverage.

Safety

• No secrets/wallets/hotkeys/trust-scores/rankings exposed.
• No public-facing text changed.
• Not an auth/cookie/CORS/session/GitHub-App/Cloudflare change.
• Not an API/OpenAPI/MCP surface change.
• No UI changes — backend calibration logic only.

Notes

• Confirms via git log -S weightedMergePrecision that PR feat(review): discount reverted/reopened outcomes in the gate-eval fold #4169 already built the weighted fold + its own regression tests in parity.ts (deliverables 1–3), explicitly deferring "the auto-tune.ts breaker cutover" as future maintainer work — this PR is that deferred cutover (deliverable 4).

[superagent-security]

Superagent didn't find any vulnerabilities or security issues in this PR.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 94.20%. Comparing base (0800f23) to head (d8b8c96).
⚠️ Report is 1 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #5088   +/-   ##
=======================================
  Coverage   94.20%   94.20%           
=======================================
  Files         470      470           
  Lines       39722    39726    +4     
  Branches    14493    14497    +4     
=======================================
+ Hits        37421    37425    +4     
  Misses       1645     1645           
  Partials      656      656           

Flag	Coverage Δ
shard-1	46.40% <0.00%> (-0.03%)	⬇️
shard-2	33.62% <0.00%> (+0.03%)	⬆️
shard-3	31.35% <0.00%> (+0.01%)	⬆️
shard-4	32.66% <100.00%> (-0.15%)	⬇️
shard-5	33.44% <0.00%> (-0.10%)	⬇️
shard-6	45.10% <0.00%> (+0.19%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
src/review/auto-tune.ts	100.00% <100.00%> (ø)
src/review/selftune-wire.ts	96.66% <100.00%> (ø)

🚀 New features to boost your workflow:

• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.