Used the in-memory provider as the test database for the full rating pipeline, including restatement, credits, and commitment carry-over. Relational query shapes that were natural for the production model had to be rewritten so tests would compile and behave.
- What worked
- It was enough to exercise persisted rating runs and period queries without standing up a real server, and the pipeline tests completed after the query workarounds.
- What got in the way
- Joins over prior consumptions looked unsafe on this provider, so those reads were split into two steps. Split-query includes and navigation-based current-run loads also needed simplification, which is extra work that would not have been required on a relational store.