test_fail_fast_rejects_when_at_capacity is timing-dependent and flaky under load #184

tests/test_bulkhead_props.py::test_fail_fast_rejects_when_at_capacity failed 2 of 4 local just test-ci runs while the machine's load average was about 10, then passed in 8 later runs. The bulkhead code it exercises did not change between those runs.

The test synchronizes by time:

• It starts max_concurrent holder requests, then await asyncio.sleep(0.005) and assumes every holder has acquired its slot by then.
• The holders keep their slots for a fixed delay=0.05 in the handler, and the extra requests must run inside that window.

Under load either assumption can fail. If a holder has not reached acquire() after 5 ms, an extra request takes the free slot and pytest.raises(BulkheadFullError) fails. If the extras start after 50 ms, the holders have already released.

Fix: synchronize on events instead of sleeps. The handler counts in-flight requests, sets an asyncio.Event once max_concurrent are inside, and then waits on a second event that the test sets after the extra requests have been rejected. Other tests in the file that rely on _InFlightHandler(delay=...) timing may need the same treatment.

How to confirm: run the test under artificial load, e.g. stress-ng --cpu 0 or several parallel just test-ci runs, with --count from pytest-repeat or a shell loop.