feat(brain): /wikilint health checker (read-only) (Phase 4)

[topcoder1]

Summary

Adds the read-only /wikilint health checker — Phase 4 of the brain wiki layer per .omc/plans/brain-2026-04-27-phases-3-5-detailed.md §4. Builds on Phase 3a/3b (PRs #29 / #32 / #33 / #36).

Surfaces four classes of issue that compound silently as KUs accumulate:

1. Near-duplicate KUs — cosine ≥ 0.95 in same (entity, topic_key), both un-superseded.
2. Temporal contradictions — overlapping [valid_from, valid_until] with conflicting text.
3. Orphan entities — entities row with <2 linked KUs and >30 days old.
4. Stale wiki pages — last_synthesis_at older than newest KU valid_from.

v1 does NO autonomous CRUD. Each finding ships with a "merge / mark-superseded / ignore" suggestion; the user runs the action manually. Top anti-pattern from both deep-research passes.

Wiring

• /wikilint slash command intercept in src/index.ts, placed before /wiki so the shared prefix doesn't route /wikilint to the entity-resolver.
• Weekly cron piggybacks the existing digest scheduler callback with a 7-day debounce via system_state.last_wikilint — no new setInterval.
• Class 1 uses qdrant.fetchKuVectors (new) — single Qdrant retrieve round-trip with with_vector: true. Pair budget capped at 500 per run; groups >32 KUs are themselves anomalous and logged + skipped.

Notes for reviewers

• Group key uses a NUL separator (\0). topic_key is space-joined by extract.ts:normalizeTopic ("current employer"), so a space separator would corrupt multi-word topics on key.split — regression test pinned in wikilint.test.ts.
• Detector functions are pure: (db, opts?) => Finding[], with findDuplicateKus async only because of the Qdrant fetch. runAll composes them.
• WikilintCommandOptions.db and WikilintCronOptions.db allow injection for testing without going through getBrainDb() singleton.
• Independent code-reviewer pass returned 1 HIGH (group-key separator), 1 MED (cron test gap), 1 LOW (comment clarity). HIGH and MED both addressed before push; LOW was a comment polish.

Test plan

• npx vitest run src/brain → 431/431 passing (+21 new tests across the two new test files)
• npx vitest run (full repo) → 2468/2468 passing
• npx tsc --noEmit clean
• npm run build clean
• Manual: run /wikilint against a real brain.db from a Telegram main-group chat — verify the report renders + the last_wikilint row appears in system_state
• Manual: deliberately trigger one issue per class (orphan + dup + stale + temporal) and confirm each section appears in the report
• Manual: wait until the next digest tick and confirm the wikilint report is delivered alongside the digest

🤖 Generated with Claude Code