Fix pre-existing test failures so backend CI turns green¶

Motivation¶

After the state_repo flat→package refactor (commit c974ace) and the parallel-test default loop (commit 5d9c569), uv run pytest -q api/tests reported 71 failures on main. The newly-added backend CI workflow (commit 5e52d08) therefore lit up red on its first push. This change makes the default loop deterministically green so the workflow blocks regressions instead of being permanently broken.

User-facing change¶

None. Test infrastructure only.

Diff summary¶

api/tests/test_state_repo.py — change the two top-of-file imports from the legacy facade api.services.state_repo to the real module api.services.state.repository. Every in-test monkeypatch.setattr uses the local state_repo binding, which now points at the module that actually owns TableClient, TableServiceClient, get_credential, _ENSURED_TABLES, JobStateRepository, JobState, etc. The legacy facade does not re-export the SDK symbols, so facade-targeted patches silently no-op'd → the eager AZURE_TABLE_ENDPOINT constructor check raised and 8 tests crashed. No production import paths are touched.
pytest.ini — switch xdist distribution from worksteal to loadfile. worksteal steals individual tests across workers mid-session; with JobStateRepository and friends now backed by per-module singletons (_DEFAULT_REPO + lock) plus daemon threads from the TestClient-driven routes, that movement causes ~12 cross-worker pollution failures and an intermittent node down: Not properly terminated worker crash. loadfile keeps every test from one file on the same worker, preserving file-level invariants while still using -n auto. Default-loop runtime is unchanged (~21-25 s on this host).

API / IaC diff¶

None.

Validation¶

uv run ruff check api — clean.
uv run pytest -q api/tests — 1369 passed in 25.41s (three consecutive full runs: 23.90 s / 23.55 s / 24.87 s — all green, no worker crashes).
uv run pytest -q api/tests -n 0 -m '' — 1435 passed in 66.09s (serial, including slow + subprocess marks).