feat(review): gate AI close on a calibrated, configurable confidence threshold

[JSONbored]

Summary

An AI consensus/split defect today carries a hardcoded confidence: 1 (consensusDefectOf / synthesizeDefect), and isConfiguredGateBlocker closes on ai_consensus_defect / ai_review_split whenever aiReviewGateMode === "block" — regardless of how confident the reviewers actually were. That is a frequent false-close source: a speculative, low-conviction AI rejection closed PRs just as hard as a certain one.

This calibrates the AI signal and gates the block on a configurable floor:

• Calibrated per-reviewer confidence (feat(docs): refresh Gittensor-aligned homepage #8): the review prompt now asks each reviewer for a confidence in [0,1] (its probability that its blockers are real defects). parseModelReview parses it with a graceful fallback of 1.0 when absent/unparseable/out-of-range, so behavior degrades to exactly today's. consensusDefectOf(a, b) carries min(a.confidence, b.confidence) (a consensus is only as strong as its weaker reviewer); synthesizeDefect([r]) / a split carry that reviewer's confidence. The confidence is threaded onto the produced ai_consensus_defect / ai_review_split advisory finding (AdvisoryFinding.confidence?).
• Threshold gate (fix(docs): add explicit favicon metadata #7): new GateCheckPolicy.aiReviewCloseConfidence (default 0.9). The AI branch of isConfiguredGateBlocker now blocks only when aiReviewGateMode === "block" AND finding.confidence >= (aiReviewCloseConfidence ?? 0.9). Below-threshold AI defects stay advisory (visible, never block). A finding with no confidence is treated as 1.0 and always clears the floor — byte-identical to today. All other blocker codes are untouched.
• Config-as-code: aiReviewCloseConfidence resolves from .gittensory.yml gate.aiReview.closeConfidence through the focus manifest into the effective settings, mirroring exactly how slopGateMinScore / qualityGateMinScore flow (parse + clamp to [0,1] without rounding, serialize/round-trip, resolveEffectiveSettings overlay). No DB migration — it stays on the config-as-code/manifest path, like those minScore thresholds. Documented in docs/review-configuration.md.

Advisory-mode-safe — no live change. aiReviewGateMode defaults to advisory (never blocks), so the live posted gate is unchanged; this only affects the BLOCK path and the dry-run "would-be" verdict.

No issue: small, self-evident hardening of an existing false-close path (the summary explains the rationale).

Scope

• The PR title follows type(scope): short summary Conventional Commit format.
• This PR is focused and does not mix unrelated backend, UI, MCP, docs, dependency, and deploy changes.
• This follows CONTRIBUTING.md and does not reintroduce GitHub Pages, VitePress, site/, or CNAME.
• I linked an issue, or this is small enough that the summary explains why an issue is not needed.

Validation

• git diff --check
• npm run typecheck
• npm run test:coverage locally — both sides of every new ?? / ternary / threshold comparison covered (ai-review.ts 99.03% stmt / 98.68% branch; advisory.ts 99.3% stmt / 98.03% branch; focus-manifest.ts 99.73% / 99.33%; the new confidence branches in advisory.ts are 100% on both sides).
• npm run test:ci (full gate, incl. actionlint / migrations / workers / mcp / ui:openapi:check / ui:lint / ui:build) — green.
• npm audit --audit-level=moderate — 0 vulnerabilities.
• npm run ui:openapi:check — no drift (config-as-code-only; no API field added).
• New behavior has unit tests for new branches, the fallback path, and the threshold boundary (both sides).

Safety

• No secrets, wallet details, hotkeys, coldkeys, user PATs, private keys, raw trust scores, private rankings, or private maintainer evidence are exposed.
• Public GitHub text stays sanitized, low-noise, and does not imply compensation guarantees or optimization tactics.
• Auth, cookie, CORS, GitHub App, Cloudflare, or session changes include negative-path tests. — N/A (no auth/session/CORS surface touched).
• API/OpenAPI/MCP behavior is updated and tested where needed. — N/A (config-as-code-only; no API/MCP surface; OpenAPI check confirms no drift).
• UI changes use live API data or real empty/error/loading states. — N/A (no UI change).
• Public docs/changelogs are updated where needed; changelogs are only edited for release-prep PRs. — docs/review-configuration.md updated; CHANGELOG.md untouched.

Notes

• No DB migration by design — aiReviewCloseConfidence is an optional, manifest-only field on RepositorySettings (it stays undefined from the DB and is set only by the .gittensory.yml overlay), keeping it on the same config-as-code path as the minScore thresholds.

[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 95.50%. Comparing base (810ddaf) to head (c8dad10).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1599   +/-   ##
=======================================
  Coverage   95.50%   95.50%           
=======================================
  Files         203      203           
  Lines       21985    22008   +23     
  Branches     7943     7956   +13     
=======================================
+ Hits        20996    21019   +23     
  Misses        413      413           
  Partials      576      576           

Files with missing lines	Coverage Δ
src/queue/processors.ts	88.25% <100.00%> (+<0.01%)	⬆️
src/rules/advisory.ts	97.60% <100.00%> (+0.04%)	⬆️
src/rules/predicted-gate.ts	100.00% <100.00%> (ø)
src/services/ai-review.ts	98.63% <100.00%> (+0.04%)	⬆️
src/signals/focus-manifest.ts	99.17% <100.00%> (+<0.01%)	⬆️

🚀 New features to boost your workflow:

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