A definition can be written by hand

[WaylandYang]

Closes #562. The capability cut; the page follows once #552 has landed (both touch Mappings.tsx).

A row in concept_mappings had exactly one way in: exploration proposes, a person confirms. A data team that already has its definitions — a metrics doc, the SQL behind last quarter's report — had nowhere to put them, and #520 measured what they are worth once they are in the prompt: 1 of 18 right answers without, 17 of 18 with, on the wide corpus. Exploration cannot fill that gap (0 of 18 there; it cannot see a convention that is not in the schema). This is the door.

What changes

• POST /kbs/{id}/mappings (Editor): concept name, kind (metric / dimension, default metric), a mounted source by name, and the same table_name / expr / sql / unit / summary / derived the revise form takes. The row lands confirmed with the author in a new written_by column (migration 0046), so a written row and an exploration-proposed one stay distinguishable — decided_by cannot tell them apart, since a confirmed proposal has one too. Audit mapping.written.
• POST /kbs/{id}/mappings/preview (Editor): render the definition (sql verbatim, else SELECT expr FROM table — the same rendering the bench scores with), run it through the read-only gate on its source, return the first row. The person sees the number before saving; the engine's error comes back verbatim when it does not run.
• Validation shared with revise (definition_shape): at least one of table, expression, SQL. The source must be mounted (source_not_mounted). One definition per (concept, source): a second is a 409, since the existing row may be a confirmed proposal and overwriting it would erase a decision — revise instead. Exploration cannot clobber a written row either: propose only refreshes proposed rows, the same rule that protects a rejection.
• ConceptMapping gains written_by; the three list queries return it.

The concept still becomes an entity of the builtin Metric / Dimension class, because that is the table's current shape. 0035 retires that shape (#556) and keeps the table, rendering into it, so a written SQL survives.

Verified

• Unit: definition_shape, render_sql (sql wins and drops a trailing ;; table alone is not runnable).
• DB test a_definition_can_be_written_by_hand: written row is confirmed with written_by; same key again is Conflict; a later propose on that key leaves it untouched; exploration's own rows have written_by = NULL.
• End to end on a fresh base with the wide corpus mounted: preview of the GMV definition returns 32931921.08, the bench truth; a broken expression returns 422 with column "no_such_col" does not exist; create → 200 and the list shows the row confirmed / written; same key → 409; unmounted source → 422 source_not_mounted; no definition → 422 empty_mapping.

Does not touch Mappings.tsx (#552) or chat.rs (#548).

🤖 Generated with Claude Code