A shape of statement implies a fact of another property, read once per phrase and cached

[WaylandYang]

The fifth slice of 0044 decision 3, and the one the prototype's recall gap points at (0044 §2): the facts a reader draws without the text stating them — a place's country from its region, a film's country from "British film", a year from a phrase — reach the typed graph through implication rules with cached readings. This is the largest remaining piece of alignment (cut 2); with it the parity run against the bound pass becomes possible once the bench in #880 is run with a model.

Model

• An implication rule is keyed like a binding (a phrase signature: phrase × subject class × object class or value) or by a kind word, concludes one property, and takes its object either from the statement's own object or from a reading of the object's words. Three readings exist, named in one table (READINGS) that feeds both the prompt and the validator: country_of_nationality, country_of_place, year_of_phrase.
• Rules have the binding lifecycle: the aligner proposes (after it has decided the signatures of a run, and once per kind word), with the same fingerprint basis as 0053 so changed inputs re-ask; "nothing implied" is recorded as an agent rejection so it is not asked again; a person approves or rejects on the alignment queue, and the decision commits with its job (0051 rule). A person's decision is never overwritten.
• Readings are model calls made once per distinct phrase by a read_phrases job and cached in phrase_readings — a name resolves to an entity in the base (found by name, or created with its name fact per 0041), a year lands as a value, and "no answer" is cached too, so an unplaceable phrase is never asked about twice. The job then queues materialisation.
• materialize never calls a model. It reads the cache and, for approved rules, writes facts marked implied with the triggering statement's evidence copied, tracked in implied_fact_sources (rule + statement, or rule + entity for kind-word rules), and retired by source exactly like bound rows: a rejected rule, a dead statement, an entity whose kind word changed all take the row down at the next run. A reading not yet cached simply produces nothing that round. A thing never implies itself.
• The export carries utopia:implied on such facts, so an auditor can tell a stated fact from a drawn one.

Schema (migration 0073): facts.implied, implication_rules, phrase_readings, implied_fact_sources. CURRENT_SCHEMA_VERSION → 72 (it counts files; 0071 is held by #861).

Surface: POST /kbs/{id}/review/alignment/rules/{rule_id} ({approve} → 202 with a job id), a rule item on the alignment queue with the shape, the concluded property and the reading, a card on the Review page in both languages.

Verified against pgvector/pgvector:pg16 on a fresh database:

cargo fmt --all --check                                     ok
cargo clippy --workspace --all-targets -- -D warnings       clean
cargo test -p utopia-store   (UTOPIA_TEST_REQUIRE_DB=1)      335 passed, 0 failed
cargo test -p utopia-server  (UTOPIA_TEST_REQUIRE_DB=1)      397 passed, 0 failed
cargo test -p utopia-extract / utopia-cli                    pass
cd web && pnpm build && pnpm test                            tsc clean, 132 passed

New coverage: a proposal lands once and a person's decision commits with the right job (read_phrases when a reading is needed, materialisation otherwise); an approved rule waits for its reading, then implies a fact with evidence and sources, and a second run changes nothing; rejecting the rule or losing the statement retires the implied row; a kind-word rule implies a fact for every thing so called and "no answer" is cached; the aligner offers a kind word and records the model's proposal on the queue; read_phrases fills the cache, names a thing and queues materialisation; the route answers 202 and scopes by base; the rule and reading parsers count malformed rows without believing them; the existing lifecycle regressions now script the rule question.

Not in this PR, said in the record: a revision table of rules; readings beyond the three; evidence for kind-word rules (the kind word is the entity's, not a statement's, so there is no quote to copy).

🤖 Generated with Claude Code