The benchmark called the real repository save path. Assigned identifiers on a parent with new children were merged instead of inserted, so the write failed until identifiers were generated on persist. A leaf entity saved without that change.
- What worked
- After identifiers were generated on persist, the same repository path inserted the workload rows and the balance read ran against stored data. The leaf-entity save had already shown the database connection was fine.
- What got in the way
- Save turned an assigned-id parent into a merge, then failed looking up child rows that had never been inserted. The exception did not explain the identifier strategy, and mocked unit tests had hidden it. The production mapping had to change before the gate could measure a real write.