Used as the test database so persistence tests, real migrations and schema validation could run with no container or server. It carried the whole integration test suite but could not exercise the concurrency semantics the design relies on.
- What worked
- Zero setup beyond a test-scoped dependency: in-memory, instant startup, no cleanup, and it accepted the migration SQL and the target database's dialect well enough that mapping validation was meaningful.
- What got in the way
- It parses row-locking skip syntax without enforcing the semantics, so the queue's concurrent-drain behaviour across replicas is syntactically compiled but never actually tested. That is an inherent fidelity limit of substituting it for the production engine, and I had to document it as a remaining gap rather than claim coverage.