semantic search: strut keeps the graph's vectors (the pass, the index reconcile, typed search through the domain indexes, the no-embed mark)

[Evanfeenstra]

Builds plans/semantic-search.md (#139), the strut side. On swarm38 a question with no keyword in it finds nothing by meaning: code nodes come from stakgraph's ingest, which never embeds, and what jarvis did embed is the bare name under a Schema index that predates its own library. This makes strut the thing that keeps the graph's vectors current, from the live :Schema meta-graph, on any Neo4j 5.

The pass — src/graph/embeddings.ts, EmbeddingPass. A timer in the serving process, started when the graph backend opens. A node of a marked type is stale when its stored Data_Bank is not the text the writer would compose today from the Schema's index fields, or its vector is missing — one comparison that catches a node never embedded, one embedded before its index widened, a field filled in or edited later, and a text stakgraph's re-sync reset. It writes what the writer writes (Data_Bank, text_embeddings, _search_fields_used, the resolver's domain label), pages by node_key (an index seek per page), strips the no-embed types last, embeds newcomers by date_added_to_graph between hourly sweeps, works 15 s a minute, and nothing waits on it. GET /graph/embeddings reports the sweep, per-type counts, and the types it skips and why. The old awaited Domain_strut backfill is gone.

The Schema index reconcile — src/graph/index-reconcile.ts, a boot step. A live index that is a non-empty strict subset of the bundled library's (Function: "name" → [name, summary, description]) is widened to the library's list — jarvis's own values, written directly, since its Schema API re-embeds a whole type inside one request and rebuilds the shared keyword index per type. Once per fixture version (a Migration ledger row), so a list jarvis narrows on purpose is never fought over. The fixture gained a version field.

Search — routeIndexes / overfetch in search.ts. A typed search reads the requested types' domain vector/fulltext indexes with an over-fetch (k from cap×10 by ×4 to 2,000 — the database cannot filter a vector query, so type is a WHERE after the nearest-k cut); an untyped one reads the global text_embeddings_vector_index, which the reader never used before (domain_all_vector_index never exists on 5.19).

The no-embed mark — embed: false on a Schema (ruling 2: Turn, StrutToolCall, StrutTurn get no vector). embed joins the core-key mirror, the resolver reads it, the writer composes Data_Bank but no vector for it and update drops a stale one, the pass strips what such nodes hold. Strut's two types carry it in code only — never flattened onto the Schema node, since an older jarvis reads an unknown Schema key as a required attribute. MiniLM runs on two intra-op threads.

Where the build differs from the plan's text (in the plan's status header): the reconcile only widens (no string→list normalization); the pass addresses a node by elementId; the newcomer tick is one range seek across all marked types; a failed page moves the sweep to the next type; a tick logs only when it did something; the per-field vector_index vectors are filled by the sweep alone; the status read has no web client.

Verified: tsc, npm test (1417), npm run test:graph (268, the full live suite). New tests: the pass end to end (another writer's node; a later description, an edit and a stakgraph-shaped reset re-embed it; a clean sweep writes nothing; unmarked types left alone and named; the cursor over 2,500 nodes with a tick resuming at it; newcomers without a sweep; no-embed on write, update and sweep; labels and hidden types; the route), the reconcile (widen / leave / stamp / no re-run under the same version), routing and over-fetch pure and live (a code type found through its domain index among 1,300 nearer decoys; a type with no domain through the global index past them), the writer's gate, the resolver's mark.

Companions, in their own repos, not stacked — strut waits on neither: stakwork/jarvis-backend#3141 (the mark as a core key, four enqueue sites gated, Turn marked) and stakwork/stakgraph#1762 (Data_Bank set once on create). Until each is deployed the pass heals what they do. After the jarvis one lands on a seed, the ontology fixture is re-dumped and one test stops setting Turn's mark by hand. Then the swarm38 validation in the plan, after the mcp pin.