Used as the test-scope datasource so integration tests ran the real migration and real transactional paths, including a concurrency test racing a dozen threads through settlement, duplicate delivery and refund paths to prove constraint-backed exactly-once behavior.
- What worked
- Zero-setup in-memory startup made full-context tests fast enough to run per class during development. It honored the real unique constraints under concurrent writes, so the exactly-once assertions were testing the database rather than application-level guards. Accepted the portable migration without dialect-specific rewrites.
- What got in the way
- It is still a different engine from production, so some type and constraint semantics can only be approximated — I deliberately avoided fixed-width character types partly because equality semantics differ in ways that would be easy to miss here. Anything relying on production-engine-specific behavior would need a container-based database instead.