Sprint 4 PR 4.9 — make the Makefile reset the DB the tests actually use

[tomqwu]

Summary

make test-all — the gate CLAUDE.md tells contributors to run before pushing a PR — has been red on main on any machine that had run the suite before. Measured at a09e422: 9 failed / 333 passed.

The cause is a path mismatch, not a flaky test.

The two halves that disagreed

• tests/conftest.py:15 sets TESTING_FORCE_MEMORY=true unconditionally, which makes api/database.py:18-19 bind to /tmp/signupflow_test.db regardless of DATABASE_URL.
• The Makefile defined TEST_DB_PATH := $(abspath test_roster.db) and rm'd that file between runs — a file the suite never reads. setup_test_data seeded it too, via the Makefile's exported DATABASE_URL.

So @echo "🔄 Rebuilding fresh SQLite test database..." was a no-op. Every run.

Why that turned into failures

The Makefile also exports SKIP_TEST_DB_FIXTURES=true, which makes conftest skip the per-test truncation fixture. With truncation skipped and the reset hitting the wrong file, nothing ever cleaned /tmp/signupflow_test.db. It accumulated across every run the repo had ever done — 518 organizations and 794 people by the time I diagnosed it. Tests that create fixed IDs (person_001, avail_test_org1, …) eventually hit rows left by earlier runs and got 409 CONFLICT. Failures grew run over run: a partially-dirty DB gave 2 failures, a dirtier one gave 9.

CI never noticed because it runs bare pytest on a fresh runner and never invokes make.

Fix

Point TEST_DB_PATH at /tmp/signupflow_test.db so both the reset and the seed target the DB the suite actually binds to, and derive every rm from that variable rather than hardcoding the stale filename.

-TEST_DB_PATH := $(abspath test_roster.db)
+TEST_DB_PATH := /tmp/signupflow_test.db

-	@rm -f test_roster.db test_roster.db-shm test_roster.db-wal
+	@rm -f $(TEST_DB_PATH) $(TEST_DB_PATH)-shm $(TEST_DB_PATH)-wal

Tests

New tests/unit/test_make_test_db_path.py (4 tests) pins both halves of the contract so they cannot drift apart again:

1. api/database.py still redirects to the /tmp path under TESTING_FORCE_MEMORY
2. conftest still sets that variable
3. the Makefile's TEST_DB_PATH equals that path
4. no Makefile recipe still references test_roster.db

Validation

Check	Result
make test-all (from a dirty DB)	unit 350, api 332, cli 16, integration 324 — zero failures (was 9 failed / 333 passed)
make test-all run twice back to back	byte-identical counts — the target is now idempotent
pytest tests/unit tests/api tests/cli tests/contract tests/web	916 passed, 21 skipped
black / ruff	clean
mypy api/utils api/core api/schemas	clean, 61 files

The two-consecutive-runs check is the one that matters: before this change the second run was strictly worse than the first, which is the signature of the accumulation.

Follow-ups

• The create-tests still use fixed IDs and are only safe because the DB is now reset between runs. Giving them unique per-test IDs would make them robust regardless of DB state — worth doing, but it's a broader change across several files and doesn't belong in this fix.

---

Generated by Claude Code