A bound statement becomes a typed fact

[WaylandYang]

0044 cut 2, second slice: the typed graph is computed from the open graph through the phrase bindings of #751 (decision 3: "typed rows carrying from_statement_id; materialisation and recomputation of the typed graph"). Pitfall 19 of docs/design/prior-work.md (a pipeline that cannot be re-run drifts).

What lands

Migration 0067: facts.from_statement_id (a typed row's source statement; checked to be typed-only; partial index), cascading with the statement.

utopia_store::materialize: one set operation per base, run at the end of every align_phrases job. Every live open statement under a bound signature becomes one layer = 'typed' row: predicate = the bound property, ends by the binding's direction (reverse swaps them), value, world-axis interval (valid_* as 0045 resolved it on the statement), attestation and confidence copied; the statement's evidence rows and role-word qualifiers are copied to the new row; from_statement_id points back. A statement with a mood qualifier is never materialised. Recomputation is the same operation: rows whose source no longer holds (the statement invalidated, the signature no longer bound, the property or direction changed, a mood added) are invalidated, rows that are due and missing are added, rows whose binding is unchanged keep their id, evidence and recorded time. Idempotent: a second run changes nothing. Ids are v7 from Rust (Postgres 17 has no uuidv7()), so rows are inserted one by one after one selecting query.

Test (a_bound_statement_becomes_a_typed_fact, database-backed, skipped without UTOPIA_DATABASE_URL): a statement under a forward binding gives one typed row with the property, copied interval, one evidence row and the since qualifier; the statement with a mood gives none; a second run is a no-op; flipping the binding to reverse retires the row and adds one with the ends swapped, the old row kept as invalidated; setting the binding to none retires it and adds nothing.

Measured

No model call. On the two bases of #751 after a run under the strict rule:

base	bound signatures	typed rows	with evidence	with role-word qualifiers	with a world-axis interval
25-document batch (#743 extraction)	36	184	184	65	3 (that base predates the time resolution of #740)
NVDA four releases (#748 extraction)	29	125	125	70	0 (table rows carry their period as time words, resolved per document; none of these carried a resolved interval)

By property on the batch: value 106, change_from_prior_period 35, votes_for / votes_against / abstentions 10 each, officer_of 6, participated_in 4, filed 2, issued_by 1. On NVDA: earnings_per_share 21, operating_income 20, net_income 18, change_from_prior_period 12, gross_margin 11, the three vote counts 10 each, revenue 5, officer_of 4, partnered_with 2, invested_in 1. A second run on each base retired 0 and added 0.

The judge of #751 already read these typed facts against their chunks (95% stated on the batch, 76% on the NVDA releases); the rows are the same statements under the same bindings, now in the ledger where the timeline, the conflict detector and the signature checks read layer = 'typed'.

Not in this change

Merging the typed rows that several statements produce for one triple (the timeline's refinement in insert_fact_inner is not applied; one row per statement for now); implication rules and readings; typed qualifiers (fact_qualifiers by declared attribute) from role words; the alignment queue (#725).

🤖 Generated with Claude Code