v0.31.0 — cooperative index yield + concurrency/self-heal fixes

[hamr0]

Summary

Ships the index-fix work as one release (0.31.0):

Feature (LC-3): index({ yield: true }) — cooperatively release the event loop between per-file parses (setImmediate) so a co-hosted timer/socket breathes during a large/force pass. Byte-identical stored index; default false unchanged. (bareloop upstream ask LC-3.)

Fixes (all validated + regression-tested, mutation-verified):

• Atomic force/self-heal rebuild — the destructive clear is folded into applyChanges' transaction, so a concurrent reader (e.g. the 0.30.0 background warm-index firing while the model recalls) never sees an empty index. Reproduced: recall returned 0 mid-rebuild; now returns the full set.
• Separate mem_embeddings table — written-memory vectors no longer share file_embeddings, where a global-tier remember() whose id equalled a file path clobbered that file's vector (and a forget deleted it). Kind-scoped getEmbeddings + one-time auto-migration on open.
• Backfill drift guard — skip a file whose disk bytes drifted from its stored hash; batch backfilled vectors in one transaction.
• MCP — retry a failed warm-index on the next initialize.

Verification

• node --test → 416 pass / 0 fail / 1 skip (pre-existing embeddings-dep skip); +6 new regression tests (LC-3 yield, concurrency ×3, embeddings ×2).
• tsc --noEmit → exit 0.
• /ship, /security, independent /diff-review on the range → all clean (0 Critical / 0 Warning).

Notes

• Pre-existing, non-blocking (flagged, not fixed here): index({ force:true, paths:[…] }) still clears the whole index while re-inserting only the scoped subset — identical to prior behavior, not reachable via CLI/MCP.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MEtvo8nTSpwN59712wEPEt