One typed fact for many statements

[WaylandYang]

0044 cut 2, third slice: several statements that produce the same triple give one typed row. Follows #752, which gave one row per statement and left the merge for later.

What lands

Migration 0068: typed_fact_sources (fact_id, statement_id), one row per statement behind a typed row, backfilled from from_statement_id; from_statement_id stays as the first source.

Materialisation goes through the typed graph's own gate. Instead of inserting a row per statement, utopia_store::materialize calls insert_fact / insert_value_fact for each due (statement, binding) pair, so the rules the rest of the typed graph already lives by apply: the same assertion with the same start reuses the row and only attests it earlier; a bare row (no time) is superseded by a dated observation of the same assertion, with the supersedes link and the evidence carried over; an observation that says "ended" closes the open row. A row's sources are then whatever statements it was computed from; its evidence rows and role-word qualifiers are the union of theirs. When a new row supersedes an old one, the old row's sources move to it in the same pass, so one run converges.

Recomputation, restated for many sources. A source is dropped when its statement no longer holds under a bound signature that yields this row (statement invalidated, signature unbound, property or direction changed, a mood added, or the row itself superseded); a computed row is retired only when it has no source left; due pairs are added as before. Idempotent: a second run changes nothing.

Test (a_bound_statement_becomes_a_typed_fact, database-backed): a bare statement, a dated one and a second document's dated one on the same pair give one live row with three sources and two evidence rows in one run (the bare row superseded, its source moved); the statement with a mood gives none; a second run is a no-op; invalidating one statement keeps the row with one source fewer; flipping both signatures to reverse retires the row and rebuilds it with the ends swapped; setting both to none retires it and adds nothing.

Measured

No model call. The two bases of #751, cleared and re-materialised under this branch:

base	statements under bound signatures	live typed rows	sources	merged into an existing row	bare rows superseded by dated ones	second run
25-document batch	184	175	184	8	1	0 / 0 / 0
NVDA four releases	125	98	125	7	20	0 / 0 / 0

The merged rows are figures a release states twice (a headline and a table: "revenue —value→ $96,221" with three sources) and the superseded ones are a table's cell first read without its period and then with it, all within one document; across documents the corpora repeat little. Before this branch (#752) the same bases held 184 and 125 rows, and a second run after the timeline's refinement had to run twice to settle.

Not in this change

Implication rules and readings; typed qualifiers from role words; the alignment queue (#725); the conflict detector's view of two sources that disagree (they stay two rows under insert_fact's rules, as they do for any other writer).

🤖 Generated with Claude Code