Problem

On every align_phrases run, phrase_alignment::shortlist embeds the text of every live "wide" signature in the base: every signature with more than SHORTLIST (10) structural candidates. It doesn't embed only the new or changed ones. It sends batches of SHORTLIST_EMBED_BATCH (32), one at a time, before the run decides anything. Nothing caches the vectors between runs, so a run that has two new signatures to decide still re-embeds every unchanged signature in the base first.

This is still the case on dev (ee3d1e2), in shortlist() in phrase_alignment.rs.

There are two consequences:

1. The per-run floor grows linearly with the base. We measured 16–19 ms per wide signature, which is the embedding round-trip with batches in series:

   • ~4.4k wide signatures: 69–103 s per run;
   • ~17k wide signatures: ~270 s per run (claim → 候选短名单开好).

   A base with ~700k wide signatures would spend about 3 h embedding on every run. That cost repeats for every small incremental batch of documents, and it's paid in embedding calls each time.

2. When the base's properties have no vectors, every one of those embeddings is discarded. nearest_relation_type_ids returns nothing and the loop continues for every signature, so the run logs shortlisted=0 after embedding the whole set. On rc7, a base's relation types only get vectors when embed_ontology runs, which happens when the workspace's embedding settings are saved. So a base created, or given its properties, after that save has no property vectors, and pays for the embeddings without getting a shortlist. Refresh the ontology vector index when the editor creates or relabels a class or property, before alignment runs #956 on dev fills vectors when the editor creates or relabels a property, which narrows this case but doesn't remove it.

Repro (generic)

1. Configure a workspace with an embedding model, then create a base and add more than 10 relation types with no domain or range. Most signatures are then "wide".
2. Ingest some documents so align_phrases runs.
3. Ingest one more short document. The next align_phrases run logs 候选短名单开好 wide=N shortlisted=0. That N is every wide signature in the base, not the few new ones. Claim → shortlist time scales with N.

Suggested fix

• Cache signature vectors. Key them by embedding model plus a hash of the embed text (phrase · example · quote), so a run only embeds signatures whose text it hasn't seen. The text of an unchanged signature is stable: its examples are the first three by recorded_at, id.
• Skip the embedding when it can't be used. If the base has no relation-type vectors of the needed kind (relation or attribute), check that once with an EXISTS and don't embed at all. That matches today's result, an empty shortlist, without the calls.
• shortlist_kind_words has the same shape, so the same treatment probably applies there.

Either change keeps the fingerprints, and therefore the decisions, unchanged. A vector cache only removes repeat work.