The semantic layer gets measured, and exploration learns to align

[WaylandYang]

Three measuring benches for the semantic layer, the coverage cut of #503, and decision record 0036.

Benches (scripts/bench/)

• mappings.mjs (A mapping proposal has never been scored #501) scores what exploration proposes. Judgement is by number: a truth entry is a definition plus one gold SQL, the proposal is run against the same rows, same number means same definition whatever the model named it. Four columns — covered, right, wrong, broken — and wrong (runs, matches nothing) is the one that decides whether automatic confirmation could ever be default-on. Two corpora: TPC-H (schema and definitions from the specification's 22 queries, rows from generate_series, comments as a separate file so their value is measurable) and wide, a flattened e-commerce order table whose every definition carries a convention the columns do not state.
• ask.mjs (A chat answer is never checked against the data #520) scores chat end to end: it asks each question, re-runs the SQL the model actually ran, and compares with the gold — two columns, sql_right and answer_right, because they come apart. --replay re-scores from the conversations already in the base (seconds, not forty minutes); --seed writes the truth as confirmed definitions (the upper bound); --confirm confirms what exploration proposed (the product path); --parallel N asks several at once.
• lib.mjs holds the one same() both benches use, so their numbers are on one scale.

Numbers, recorded on #501 and #520:

	exploration covered	chat right
TPC-H, none	8/24 after #503 (6/24 before)	23/24
TPC-H, exploration confirmed	—	24/24
wide, none	0/18	2/18
wide, conventions as one page of prose	—	14/18
wide, truth seeded	—	17/18

#503 — an exploration says what it covered

proposal_cap scales with the schema instead of a fixed 12 (TPC-H coverage 25% → 33%); the schema-text cap is now a real cross-source bound; a run writes a row in mapping_exploration_runs (tables and columns scanned, truncation, cap, returned, accepted, drops with an example each, tables covered, error); last_run on GET /kbs/{id}/mappings. Single-source bases accept whatever source the model echoes — the first bench run lost all twelve proposals to a source name with a timestamp in it. Migration 0044.

0036 — Exploration aligns a schema to the ontology

Exploration today inserts two classes and makes concepts entities of them; the wide base had 93 concept entities, most of them column names. The record moves the concept: an attribute of a real class, or a rule over such attributes; exploration proposes an alignment per table through ontology_proposals; conversions are 0032's expression tree; concept_mappings becomes rendered. Revision notes on 0009 and 0011. Decision 7 is implemented in #561; the written-definition door is #563; the conventions-as-prose dead end was measured at 14/18 and revised in place. Follow-ups: #554, #555, #556.

Also renumbers the exploration-ledger migration to 0044 (dev took 0041) and merges dev.

🤖 Generated with Claude Code

[WaylandYang]

Review before merge. Two ledger gaps fixed, one bench guard, one renumber:

• The cap was invisible. Proposals past cap were sliced off before the loop with no drop reason, so returned − accepted ≠ Σ dropped whenever the model returned more than the cap, which is exactly the wide-table case the ledger exists for. The excess is now recorded under cap with the count and the cap value.
• A decided definition counted as accepted. propose upserts only WHERE status = 'proposed', but the loop counted every call. A second exploration on a base with confirmed or rejected definitions reported accepted = N while Review gained nothing. propose now returns whether it wrote (rows_affected), and a proposal that lands on a decision is recorded under decided. Coverage still counts the table: it has a definition, which is what coverage asks. Test a_rejected_mapping_does_not_come_back asserts the flag.
• finish failing is logged instead of swallowed.
• lib.mjs: BENCH_PSQL without -d used to make every onDb() fall through to the command's default database, DROP SCHEMA included; -d is now inserted when absent.
• The decision record is 0036: dev took 0035 for the vector index (A vector index is built by a job #566) while this was open. README rows read 0032 … 0036 in order once this merges.

Not changed: migrations here use plain CREATE TABLE, like 0038–0042 on dev; CI's second sqlx migrate run skips applied versions, so IF NOT EXISTS only matters for a renumber after a base has run the old number, which 0044 has not.