Python suite takes ~24 min serially; no parallelism, and per-test app setup costs up to ~10 s #524

Tier 3: feedback loop

Measured

make test-py on main (ec46dc4) in a cloud dev container:

4470 passed, 11 skipped, 65 deselected, 1 xfailed, 15 warnings in 1452.50s (0:24:12)

• Neither pyproject.toml, the Makefile nor pr.yml uses pytest-xdist, so the suite runs on a single core.
• The slowest entries are mostly fixture setup, not test bodies:
  • tenants/tests/test_resolver_source.py::test_header_source_and_vary: 9.8 s setup
  • api_tokens/tests/test_api_tokens_service_keys.py::...: 9.8 s setup
  • branding/tests/test_tenant_branding_cost.py::...: 9.2 s setup
• Those setups build a fresh create_app with its lifespan, and create every module's tables, for each test.

Cost

A 24-minute local run means contributors stop running the full suite and rely on CI. That defeats the "prove it before pushing" discipline the project asks for.

Fix

1. Add pytest-xdist and run -n auto in make test-py and CI. Fixtures already use per-test in-memory SQLite, and Postgres resets its schema per test, so isolation should mostly hold. Audit any process-globals first, for example Audit capture is a process-global; writes made in Celery tasks are never audited #498's _db_state.
2. Profile create_app and table creation under the app / db_session fixtures. Candidates:
   • Cache discovered modules and built metadata per session.
   • Create the schema once per worker and use a transaction or savepoint per test where tests don't need a fresh app.
3. Track total suite time in CI, so regressions get noticed.