Summary

test_service_info_async_request flakes on multiple CI runners (e.g. macOS 3.12 skip_cython, ubuntu 3.13 use_cython) with:

AssertionError: assert [b'\n\x00\x01\x02'] == [b'\n\x00\x01\x03']

Root cause

After tracing the captured packet log, the failure isn't a timeout / convergence problem — a stale queued multicast response ends up overwriting the correct one:

1. The earlier async_get_service_info / gather calls trigger queries that the local zeroconf answers via out_queue (or out_delay_queue). The queue snapshots the records (SRV → ash-1.local., A → 10.0.1.2) and aggregates the send by up to ~620ms / ~1320ms.
2. async_update_service(new_info) swaps the registry entry to ash-2.local. / 10.0.1.3 and broadcasts the new records.
3. With quick_timing, the broadcast finishes within ~30ms, so the stale queued response is still pending when the test reaches the assertion block.
4. The test calls _clear_cache(...) and async_request(zc, 500) with _is_complete patched to False.
5. The queued response flushes after the cache clear, re-caching ash-1.local. → 10.0.1.2 and pointing aiosinfo.server at ash-1.local..
6. The patched loop then keeps asking for an A record for ash-1.local., which nothing in the registry answers, so the final state stays on 10.0.1.2 and the assertion fails.

Evidence from the failing ubuntu run (MulticastOutgoingQueue flushing stale records right after the ash-2 broadcasts):

... answers=[(record[srv,…,xxxyyy._test1-srvc-type._tcp.local.]=120/119,ash-1.local.:80, 0.0),
            (record[txt,…]…, 0.0)],
   additionals=[record[a,…,ash-1.local.]=120/119,10.0.1.2, …]

…followed by repeated questions=[a[…,ash-1.local.], quada[…,ash-1.local.]] queries from the AsyncServiceInfo loop.

Fix

Drop the pending entries from both out_queue and out_delay_queue before clearing the cache. The already-scheduled loop.call_at for each queue fires on the empty deque and is a no-op (the while len(self.queue)… loop in async_ready doesn't iterate, and if answers: is False), so this is deterministic and adds no wall-clock delay.

A code-side fix for the underlying behaviour (invalidating queued responses when the registry changes) is left as a separate concern.

Test plan

• SKIP_CYTHON=1 poetry run pytest --timeout=60 -q tests/test_asyncio.py::test_service_info_async_request — 5 consecutive runs pass (~4.15s each).
• poetry run pytest --timeout=60 -q tests/test_asyncio.py::test_service_info_async_request (with Cython extensions) — 3 consecutive runs pass.
• poetry run pre-commit run --files tests/test_asyncio.py — clean.