kit 3.2.0

Added

• sandstream-kit-plugin-sentrux — architecture-decay findings. A read-only
  scanner plugin (same shape as the Snyk / Wiz plugins) that ingests
  sentrux check|gate --json — an architecture health score (0–10000 from
  modularity / acyclicity / depth / equality / redundancy), the baseline gate,
  and rule violations — and appends them to .kit-scan-results.jsonl so
  kit check --security can gate on architectural regressions, an axis kit
  didn't previously cover. Tolerant parser (field-name variants degrade
  gracefully); a failed gate with no discrete violations still emits one finding
  so it never passes silently. Deterministic, zero-LLM, nothing leaves; the
  operator runs Sentrux in their env and opts the plugin in via kitPlugins.
• kit memory pal claim — atomic take for parallel agents. When several
  agents share a PAL (a durable device + ephemeral cloud sessions), two could
  both start the same open item. pal claim <id> flips it to claimed via an
  UPDATE … WHERE status='open' guard, so exactly one agent wins the race
  (✓ claimed vs a no-op) and the item drops out of everyone's open list;
  claimed_by records the winner (defaults to this device). pal release <id>
  returns an abandoned claim to open. Deterministic, zero-LLM, local; schema
  v6 (older rows have NULL claim fields — backward-compatible).
• kit memory learn — surface instructions you keep re-typing. Deterministically
  mines the local store for user messages repeated 3+ times (verbatim after casing /
  punctuation normalization), ranked by distinct sessions then count, with a
  correction flag for redirections ("no", "stop", "instead", "nej", "istället").
  These recurring asks are candidates for a memory rule — record them with
  kit memory share or in a rules file (CLAUDE.md / AGENTS.md) instead of re-typing.
  Zero-LLM, local, no ML — kit finds the pattern; you decide the rule. --json
  supported. Idea from headroom's learn (kit-research), done the kit way.

Fixed

• kit memory search no longer comes back empty for multi-term queries. A
  query whose terms don't all co-occur in a single message used to match zero
  rows (the FTS5 expression joined terms by implicit AND). It now falls back to
  OR when the strict AND finds nothing, bm25-ranked so the message covering the
  most terms ranks first — relevance instead of all-or-nothing. Single-term
  lookups and exact multi-term matches are unchanged; still zero-LLM, local,
  no new deps. (#164)

---

Full changelog: https://github.com/sandstream/kit/blob/v3.2.0/CHANGELOG.md

Verify this release:

git tag -v v3.2.0
npm audit signatures