fix(test): /ws/stats tests slept 50 ms and hoped — poll the condition instead

[wshallwshall]

main is red. test (windows-2022, py3.14) failed at be2fc08c — the squash-merge of #56, whose own head was 33/33 green:

assert harness.frames, "expected at least one stats frame before revocation"
AssertionError: assert []

Not the merged change: #56 touched TLS key-exchange policy, an OIDC settings validator and docs, none of which this route reaches. The test is defective — and deliberately not treated as a flake to re-run, because this repo has already mistaken a livelock for one and no rerun or timeout ever fixed that.

The defect

The first frame cannot be sent until store.stats() returns — the route awaits it, builds the frame, then sends. The fixed 50 ms sleep had to cover that query plus scheduling the harness task, the handshake, and the handshake-time authorize. None of which these tests measure. It passes locally 3/3, in 10.06s / 12.33s / 3.89s — and that spread is the tell.

Replaced with _wait_for_first_frame, a bounded poll on the condition actually being asserted. This does not weaken the precondition: a frame is still required, and still required before the revocation. It also separates a dead task from a slow one (task.result() re-raises the real failure), because a route that blew up would otherwise look identical to a loaded runner.

A second, quieter defect

test_disabled_account_is_closed also slept 50 ms, and asserts only close_code == 1008 — which both the mid-stream revalidation and the pre-first-send re-check produce. So on a slow box the account was disabled before the first send, the close came from the wrong path, the assertion still held, and the test silently stopped covering what its comment claims while duplicating test_revoke_before_first_send_yields_no_frames. A coverage move with no symptom. It now waits for a frame, so "mid-stream" is a fact.

Making the CI failure reproducible

test_a_slow_first_stats_build_does_not_break_the_precondition stalls store.stats() to 0.4s — 8× the old budget — so the old form fails on any machine. That is the property the original failure never had: it only reproduced on a loaded Windows runner, which is not something you can iterate against.

mutation	result
M1 restore await asyncio.sleep(0.05)	slow-stats regression test RED (1 failed)
M1, second half	test_disabled_account_is_closed stays GREEN — the silent coverage move, demonstrated rather than argued
M2 _FIRST_FRAME_TIMEOUT = 0.0	3 failed — so the wait is load-bearing, not decorative

Timeouts

First-frame wait 10s; harness budget raised 5s → 20s, because the harness must outlive the wait plus the close it then waits for, or it times out mid-wait and the failure points at the route instead of the clock. Both sit inside the 60s pytest timeout, which stays the real backstop against a hang.

Verification

ruff check · ruff format --check · mypy --strict · full suite 9636 passed, 719 skipped. The one failure is the pre-existing environmental test_installed_metadata_matches_dunder_version (editable metadata 0.3.0 vs __version__ 0.3.2; its own comment names the case and the remedy).