mocking-stubbinglisted

# Mocking Stubbing The "tests must pass against synthetic data" gate. ## Mock layers | Layer | What's mocked | How | |---|---|---| | DB | All persistence | sqlite-in-memory or testcontainers + synthetic fixtures | | External APIs | Hospital HIS / 3rd party | responses library / wiremock | | LLM | Model calls | deterministic stub via mcp-model-router test mode | | Time | clocks | freezegun / fake clock | | Storage | Object store | localstack / minio | ## Discipline - **No real service contacted during Step 9**. If a test "needs prod", the test is wrong. - **No real PHI in fixtures** (enforced upstream by `test-data-generation`). - **Deterministic** — LLM stubs return canned responses; flaky tests are bugs. ## Workflow 1. Load fixtures from `mock/阶段N-*/`. 2. Stand up mock services (script: `scripts/test/up_mocks.sh`). 3. Run test pyramid: unit → integration → e2e. 4. Collect results as JUnit XML. 5. If green → emit TEST_REPORT.md + hand off to Step 10. 6. If red → hand off to `systematic-debugging`. ## Common failure modes 1. **"Just this once" real call** — test reaches real LLM. Mitigation: hooks block production endpoints in test env. 2. **Flaky LLM stub** — random response in stub. Mitigation: deterministic by seed. 3. **Fixture drift** — fixtures stale vs schema. Mitigation: schema-driven generation. 4. **Mock leak** — test uses real env var. Mitigation: explicit env clear before each test.