Our unit tests only inspect SQL *fragments*. That was not enough: the VM smoke suite found three bugs the unit tests missed (see `docs/AI-USAGE.md`, 2026-09-23). - [ ] a test database (or transactional fixtures) for the search builders - [ ] assert the decade/genre/country filters return the right rows - [ ] assert keyset paging is stable when the sort key is NULL (unrated albums) - [ ] assert facet counts match what the filters would actually return **Acceptance:** the tests run in CI without a manual setup step.