test(sync,e2e): replace busy-poll loops with wall-clock waits

[Xander-git]

Summary

Fixes the timing-sensitive test flakes that were failing CI under load (red on main, surfaced by the checks on #23). These are test-only changes — no production code is touched. Implements options B (event-driven/time-bounded waits) and C (Playwright auto-retry) from the follow-up plan.

Root cause

The sync tests drive a background worker thread and then busy-poll a fixed iteration count for a state change:

for _ in range(400):                       # ~4s budget (400 × 10ms)
    row = await client._queue.get_by_id(handle.job_id)
    if row is not None and row.state is SyncJobState.FAILED:
        break
    await asyncio.sleep(0.01)
else:
    pytest.fail("expected terminal FAILED on second hash mismatch")

The worker runs at worker_poll_interval_s=0.005. On a loaded runner the worker thread doesn't get scheduled often enough to finish its retry/reconcile cycle within the fixed budget, so the loop exhausts and fails despite correct behavior. The e2e test had the same class of bug: it read row text right after wait_for(visible), catching the transient 'Active' before the override badge re-rendered to 'Override active'.

Changes (B — sync waits)

Two shared helpers added to tests/unit/sync/_helpers.py:

• wait_until(predicate, *, timeout=30, interval=0.02, message=...) — generic wall-clock-bounded await on an async predicate.
• wait_for_job_state(client, job_id, target, *, timeout=30) — the common queue-state shape; returns the matching row, and on timeout raises naming the last observed state (strictly more informative than the old pytest.fail strings).

The budget is now a generous time ceiling (30s — irrelevant when the worker is responsive, robust when it's slow) instead of a magic iteration count. 23 busy-poll loops were replaced across:

• tests/unit/sync/test_nas_client.py (6 — incl. deleting the duplicate local _wait_for_state helper)
• tests/unit/sync/test_nas_client_extra.py (8)
• tests/integration/test_nas_sync.py (9 — queue-state loops → wait_for_job_state; filesystem/sync_state predicates → wait_until)

Negative assertions and post-loop property checks were preserved verbatim; only the initial busy-wait was swapped.

Changes (C — e2e)

tests/e2e/test_flow_06_problems.py: replaced wait_for(visible) + inner_text() assertions with Playwright's auto-retrying expect(...).to_contain_text(..., timeout=5_000).

Test plan

• ✅ ruff check + ruff format --check clean on all 5 files
• ✅ pytest tests/unit/sync -p no:randomly — 116 passed, run 3× (the previously-flaky tests, now stable)
• ✅ pytest tests/integration/test_nas_sync.py -p no:randomly — 11 passed, run 2×
• ✅ e2e file parses (full e2e run happens in CI with chromium)
• ✅ No src/ changes

Together with #23 this should bring the qc workflow fully green on main.

🤖 Generated with Claude Code