Flaky: test_voice_note_is_transcribed_and_answered misses the sendVoice reply under load #546

Test: tests/channels/test_telegram.py::test_voice_note_is_transcribed_and_answered

Where: a local full-suite run (Windows, pytest -n 8, worker gw2) while testing #542. It was not a CI run. The heavy dependency bumps don't look related: the test passed 3 times out of 3 when run alone in the same venv.

Error:

    voices = [kwargs for method, kwargs in sent if method == "sendVoice"]
>   assert len(voices) == 1
E   assert 0 == 1
tests\channels\test_telegram.py:385: AssertionError

Every assertion before it passed: the transcription, the "Heard:" echo, the model call and the text reply "It's sunny.". Only the voice reply was missing.

Likely cause: the spoken reply (voice_replies = True) is sent after the text reply, and the test asserts right after tg.say(...) returns. On a loaded machine the sendVoice call (synthesis or encoding of the reply audio) hadn't landed yet. A fix would wait for the sendVoice call with a bounded poll instead of asserting at once.