takes.embedding had readers (searchTakesVector, think's takes_vec arm,
getTakeEmbeddings) and stale-tracking (countStaleTakes/listStaleTakes)
but no writer anywhere in src/ — the vector takes arm was structurally
dead on every install.
- BrainEngine.updateTakeEmbeddings (postgres + pglite, lockstep): writes
embedding + embedded_at for active takes; updated_at untouched
(embeddings are derived data, not a content change).
- embedStaleTakes in src/commands/embed.ts: claim-text backfill wired
into the --stale path, so CLI `embed --stale`, the cycle/dream embed
phase, and sync auto-embed all populate takes.embedding. Log-and-skip
semantics; dry-run counts into would_embed; EmbedResult.takes_embedded
reports the count.
- Migration v125 (takes_embedding_dim_align): the takes DDL hardcoded
VECTOR(1536); brains configured with a different-dim embedder could
never take writes. Since the column was writer-less it is all-NULL
everywhere, so a guarded DROP/ADD retype to the configured dims (same
config source as the facts migration) is lossless and works on PGLite
too (which can't ALTER COLUMN TYPE vector(N)). No-op when dims match
or data exists.
- test/embed-stale-takes.test.ts: fails without the writer (backfill,
dry-run, idempotence, searchTakesVector goes live, v125 retype).
Fixes#2089
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
