We test heavily at the integration boundary — the places where assumptions break. Pure unit tests are useful but not the priority; we want tests that catch real bugs in real flows.
For anything that touches PHI or clinical output, we have a separate suite that runs against deidentified production-shaped data. That suite is required green before merge.
Flaky tests get one chance. After that, they're deleted or fixed — we don't tolerate "just rerun it" as a workflow.
