test_disconnected_raw_transcription_cancels_its_sidecar passes in its file and fails on its own, on any machine that has not downloaded an STT model:

tests/test_stt_transcription_cancellation.py:190
    assert raised.value.status_code == 499
E   assert 409 == 499
E    +  where 409 = HTTPException(status_code=409, detail="STT model 'small' is not
       downloaded. Download it in Settings, then Voice, before loading it.").status_code

Cause

_transcribe_audio_result does an implicit load through the registry before it ever reaches the sidecar:

load_stt, _ = _stt_lifecycle()
await asyncio.to_thread(load_stt, model, serving_engine, cancel_event)

and the registry refuses unless a complete snapshot is on disk:

snapshot_path = _find_complete_cached_snapshot(model_id)
if snapshot_path is None:
    raise SttModelNotDownloadedError(...)

The test stubs _stt_sidecar_for and _resolve_serving_stt_engine, but not the load. So the disconnect it exists to test is never reached. It gets there only when an earlier test in the file has left an engine resident, or when the host happens to have the model cached.

The sibling test directly above it already stubs _stt_lifecycle for exactly this reason:

monkeypatch.setattr(inference_route, "_stt_sidecar_for", lambda _engine: sidecar)
monkeypatch.setattr(inference_route, "_stt_lifecycle", lambda: (load, None))

Fix

Do the same. Cancellation, not download state, is the subject of this test.

Verification

Forcing the no-snapshot condition (_find_complete_cached_snapshot returning None, which is what a runner that has never downloaded an STT model looks like):

The gap between rows one and two is the order dependency.

Which PR, and which side is wrong

Both this test and the implicit registry load it trips over came from the same PR, #7984 (d20db3f1f, 2026-08-11). The test was incomplete from birth, and the sibling test directly above it, added by that same PR, does stub _stt_lifecycle.

Worth flagging: #7984 is also the PR behind the stt_unload 500 in #9013, and behind the two test_stt_ggml_sidecar.py unload tests that fail standalone on main. Three defects in one area, each invisible because a serial run happened to leave the right state behind.

Robustification

The stub now records what it was asked to load, and the test asserts it. Otherwise, if the route ever stops loading through the registry, the stub silently becomes dead code and the test stays green for a path it no longer covers. Mutating the route to skip the load:

AssertionError: the transcribe path did not load through the registry: []

Found while measuring whether the backend suite can run under pytest-xdist. --dist load splits the file across workers, so the test lands on a worker where nothing loaded an engine first and it fails. The dependency is on main today though, which is why this is a standalone fix.