The schema document is searched, not extracted

[WaylandYang]

Closes #553. Implements decision 7 of 0035 (the record itself arrives with the bench branch; its status line gets this PR's number there).

Mounting a data source ingests the schema as a markdown document so chat can search_chunks for a table. That document then went through extraction like prose, and the extractor filed every column name as an entity — on the wide bench corpus, 28 of 40 concept entities were column names.

What changes

• sources.config.extract = false marks a source whose documents are searched and never extracted. Source::extracts() reads it (default true; a non-boolean counts as unset, so a typo cannot silently stop a source). sync_schema_doc creates the "Data schemas" folder with it; migration 0045 backfills existing folders and adds skipped to graph_status.
• pipeline::process_document stops after indexing and embedding for such a document and marks it skipped — none would read as "not yet queued" in the Library, and it never will be.
• documents::queue_extraction filters these documents in SQL, so full rebuild, per-source re-extract and any future caller all skip them in one place. The two manual entry points (POST /documents/{id}/extract, POST /kbs/{id}/sources/{sid}/re-extract) refuse with source_not_extracted so a person hears why rather than "queued 0".
• Library: a Not extracted / 不抽取 label for the status; the per-source re-extract button is hidden when the source does not extract (the per-document button's status whitelist already excluded it).

Not a name rule ("folders called Data schemas never extract"): the flag is on the source, which is the thing that has the property.

Verified

• Unit: Source::extracts over {}, {path}, true, "no", 0, false.
• DB test a_schema_document_is_searched_not_extracted: full-base and per-source queue_extraction leave the schema document out.
• End to end on the bench base: fresh KB, mount wide → wide-schema.md is ready / skipped, 0 entities, 0 extract jobs; both manual endpoints return 422 source_not_extracted and queue nothing.
• Migration applied on the bench base: both existing "Data schemas" folders backfilled, CHECK widened.

Touches the ingest path #526 also wants to change; separable — this decides whether a document is extracted, #526 decides when. Does not touch Mappings.tsx (#552) or chat.rs (#548).

🤖 Generated with Claude Code

[WaylandYang]

Bench after this change, wide corpus, same scripts and truth as #501 / #520, fresh base each time.

	before (#501 run, base 01a08375)	after #553 (base 01a08642)
concept entities on the graph	93 (mostly column names, accumulated over runs)	12, all from exploration
schema document	extracted	skipped
exploration: metrics right / wrong	0 / 5	0 / 6
exploration: definitions covered	0/18	0/18
chat, no confirmed definitions	1/18 (6%)	2/18 (11%)

The graph number is the one this PR is for, and it moved from 93 to 12. The accuracy numbers did not move, and were not expected to: chat writes SQL from the schema document, which is unchanged, and exploration still cannot see a convention that is not in the schema. The one extra right answer (test_orders) is inside single-run noise. The lever for those two rows is #554–#556; the seeded upper bound on the same corpus is 17/18 (#520).