Settle derived conclusions after source cleanup removes their evidence

[Maya-Kid]

Source cleanup removed missing documents without settling the conclusions supported by them. Reuse the single-document deletion path's settle_derivations once after a nonempty batch, respecting the knowledge base's existing inference setting.

The authenticated PostgreSQL regression covers both deletion paths. It verifies that single deletion preserves cup-2's conclusion and that a pre-existing, unrelated cup-3 keeps exactly the same conclusion row IDs through both single deletion and batch cleanup. Source fixtures use api: external keys under the existing api source kind.

Validation: a8d548a, based on dev@9f50e3c: Rust format, strict workspace/all-target Clippy and the PostgreSQL 16/pgvector workspace suite passed (1,079 passed, 0 failed, 5 existing ignored). UTOPIA_TEST_REQUIRE_DB=1 was set and the cleanup regression was verified by name in the executed-test log. The unavailable PDF fallback utility limits that unrelated coverage.

The cleanup test module is registered beside the handler, avoiding the competing module insertion at the bottom of sources_routes.rs; a merge-tree check with #884 at 8e60274 succeeds. This PR remains based directly on dev, without a dependency on #884. No schema, API or scheduling policy change.

Refs #875.

[WaylandYang]

Minimal and correct: it reuses the same settle_derivations the single-delete and ingest paths already call, once per batch, gated by the base's inference setting. The real-database test walks both deletion paths on one fixture and pins the unrelated conclusion's row identity, which is the right invariant. Only note: like single delete, a failure in the settle step returns 500 after the documents are already gone; consistent with existing behaviour, so not blocking. LGTM.