BrainLayer Writer Arbitration

BrainLayer uses a single-writer arbitration path for background producers that can otherwise fight over SQLite's write lock.

Operator Contract

BRAINLAYER_ARBITRATED=1 makes producers enqueue writes instead of writing directly to the database. The current launchd templates set this for watch and enrichment.
BRAINLAYER_DRAIN_EMBED=0 disables post-drain embedding, mainly for tests or emergency operator debugging. Production should leave it enabled so queued brain_store chunks reach semantic search.
The unified durable queue lives at ~/.brainlayer/queue/. Each event is one JSONL file named by source, timestamp, and UUID.
com.brainlayer.drain.plist runs scripts/drain_daemon.py, which drains the queue every 500ms under BEGIN IMMEDIATE.
The drain daemon opens SQLite with APSW, loads sqlite-vec, writes chunks/enrichment updates, and embeds queued brain_store chunks into chunk_vectors/chunk_vectors_binary before removing their queue files.
Legacy pending-stores.jsonl is migrated by brainlayer flush; migration assigns stable chunk IDs so rerunning after a crash is INSERT OR IGNORE safe.

Startup no longer performs large synchronous trigram repairs.
Run brainlayer repair-fts for an explicit chunks_fts_trigram rebuild.
scripts/launchd/com.brainlayer.repair-fts.plist schedules that repair weekly.
Set BRAINLAYER_REPAIR=1 only for an operator-controlled process that should run the repair during VectorStore initialization.