A kind word binds to a class

[WaylandYang]

Stacked on #740. The first slice of alignment (0044 decisions 3 and 4): a kind word binds to a class, once per base, and every entity of that kind gets the class. Extraction keeps writing only the document's kind word; classes stay the ontology's; the binding is the view between them.

What lands

Ledger (migration 0065): type_bindings (base, normalised kind word, spellings seen, class or none, status bound | none | undecided, the two votes, decided_at, decided_by agent | person, unique per base and word); entity_types.updated_at so a binding can be told stale; type_source = 'aligned'. Store: signatures (each kind word with count, spellings, example names and the relation phrases its things take part in), stale (bound to a class edited since, or none/undecided decided before a newer class), decide (a person's decision is never overwritten by the agent), apply / unapply (only rows a person has not typed), bound_map, propose (a kind word with no class is offered through the existing proposed-type flow).

Contract (utopia_extract::align): one call over a batch of kind words with candidate classes and definitions; choose a class only when every thing of this kind is an instance of it by its definition, else null.

Job align_types (per base; enqueued after each document's extraction and after a class is created or edited; one run per base at a time under an advisory lock): candidates by ontology embedding when the base has one, else the whole class list when it is small; two votes per word with the candidate order reversed on the second; agreement binds or records none, disagreement records undecided for the alignment queue (#725). Bound classes are written to the entities; run_open passes a bound class into identity resolution so namesakes of different kinds separate again.

Nothing is dropped for a word with no class: the entity keeps its kind word and its statements in the open graph; the only extra is a passive "add a class" suggestion on the ontology page.

Measured on the 25-document batch (14 hand-seeded classes, no embedding model)

180 kind words: 81 bound, 94 none, 5 undecided (two votes disagreed: builders, business, business segment, department, people). 338 of 802 live entities typed (42%); the untyped ones are mostly described things whose word names no kind ("capacity", "changes", "impact"). Bindings read well across domains: company / agency / government → organization; act / reports / application → document; asset / debt / line item / cash flow → financial_metric; disease / condition / side effect → medical_condition; study / trials → clinical_trial; 指标 → statistic; 汽车 / 工业产品 → product; patients / director nominee → person. One doubtful: warning → document.

The first run, triggered fourteen times by fourteen class creations, ran fourteen jobs in parallel and hit the model endpoint with 502s; the lock fixes that, and the serialised run completed with no failures.

Not in this change

Splitting long descriptions into a short head plus the statements folded into them; binding relation phrases to properties and materialising typed facts (the next slices of cut 2); the alignment review queue cards (#725).

🤖 Generated with Claude Code