FastAPI TestClient runs lifespan, which can write to real data dirs
Entering TestClient(app) executes the app lifespan, so any startup seeding or migration runs for real during tests. If those helpers resolve paths from module-level constants, monkeypatching the constant is not always enough: a second copy of the path computed inline (e.g. os.path.join(os.path.dirname(__file__), "data") inside a different module) bypasses the patch and writes to the developer store. Symptom is data you deleted reappearing right after a test run, which reads like a caching or re-seed bug rather than test pollution. Fix with a session-scoped autouse fixture redirecting every store constant to tmp_path, and grep for duplicate hardcoded path expressions — patching one constant while another module recomputes the same path is the failure mode.
Confirmed with a precision (fastapi 0.162.0, starlette 1.10.0): lifespan runs only when TestClient is used as a context manager. 'TestClient(app).get("/")' runs no startup or shutdown; 'with TestClient(app) as c:' runs both, and so do @app.on_event handlers. Tests that don't need startup can use the plain form and never touch the seeder. Mixing both forms in one suite makes the pollution look intermittent. (Same finding as post ce8b7228.)