v1.2.0 — local-mode retrieval fusion

Changed

• Local-mode brain_search now runs the full retrieval stack. After rebundling the in-process
  engine against the INTKB upgrades, local search runs native-FTS5 + qmd reciprocal-rank fusion
  inside adapter.query(), then replicates the API's freshness/category rerank in
  src/local-server.ts (read-only store lookup for {category, updatedAt}, fused scores normalized to
  [0,1] before the freshness multiply, raw-order fallback if the store can't open). Local mode stays
  network-free while matching the team path's ranking; two 2026-07-16 incident queries that returned 0
  hits now return the correct governed memory with a qmd:// citation. (#48)

Added

• Per-platform install instructions. The README and onboarding/README.md now spell out that the
  plugin runs in Claude Code or Cowork (same /plugin install), takes a manual mcpServers
  config on Claude Desktop, and cannot run in claude.ai in a browser or the phone apps — plus a
  Windows path (/plugin commands + %USERPROFILE%\.teamkb\team.json) alongside the macOS
  one-click installer.

Fixed

• SessionEnd multi-learning slots + hook wiring. deriveCandidateId now takes
  learningIndex (0..4): key = (tenant, sessionId, session-end, index) so up to 5 learnings
  per session stay distinct while re-distill of the same index collapses. Autocapture hook
  resolves session_id / sessionId / transcript-path fallback and requires
  brain_capture({ sessionId, learningIndex }) in the distiller prompt. Client no longer
  invents already_exists from bare HTTP 200 without intake. Skill docs corrected (inbox
  does dedupe at intake).
• Session-stable capture idempotency + frozen outbox contract (Property 1/2 seam).
  deriveCandidateId uses session keys when sessionId is provided; without sessionId keeps
  content-hash UUIDv5 for manual /brain-save. Outbox freezes final POST body; drain replays
  file bytes only. Surfaces intake / alreadyExists from the server.
• /brain now retries with keywords when a full-sentence question returns nothing. Retrieval is
  keyword-AND, so a natural question ("what did the team ship this week?") could return zero even when
  the topic is well covered — a new user running the suggested proof query saw an empty result and
  assumed the setup was broken. The /brain skill now drops filler/question words and retries once
  with the strong keywords before reporting empty, and the onboarding proof query is now a keyword
  query (shipped this week) in the README and the macOS installer. (The deeper retrieval-side
  OR-fallback is tracked separately.)
• README + GitHub description corrected for rebundled local retrieval and the two-mode network
  story. The brain_search tool-surface row now states local search runs native-FTS5 + qmd RRF
  fusion then a freshness/category rerank (it was silent on both); the stale governed-second-brain
  umbrella links point to bobs-big-brain-umbrella; and the blanket "No daemon, no network"
  description becomes "local needs no daemon or network; team connects to the one brain over your
  tailnet" (the old line mis-described team mode). (#49)

Added (internal)

• Advisory MiniMax-M3 two-lane PR reviewer (.github/workflows/minimax-review.yml + REVIEW.md).
  A defect lane (secret/credential leaks, mode-boundary breaks, rebundle drift, honesty-word
  violations) and an adversarial claims lane (audits the PR description + changed docs against the
  actual diff). Advisory only — never a required check; uses the fork-safe pull_request trigger with
  a same-repo guard so forked PRs never receive the API key. (#47)