tests: _patched_pipeline leaks a Mock onto transformers.pipeline

[aviv1ron1]

Where it lives: main (latent) and feature/granite-turboctc-default (manifests). Found at ef43e69.

tests/unit/test_asr.py::_patched_pipeline does not restore transformers.pipeline. It nests two patches:

with (
    mock.patch("transformers.pipelines.pipeline", factory),
    mock.patch("transformers.pipeline", factory),
):

transformers is a _LazyModule, so transformers.pipeline resolves through __getattr__, which reads it from transformers.pipelines while the outer patch is still active and caches the Mock. On exit the inner patch restores the Mock rather than the real function.

Reproduced locally (transformers 5.8.1):

in __dict__ after: False | is the mock: False
later `from transformers import pipeline` -> mock: True

tests/unit/ runs before the integration tiers, so every later real ASR load in the same session gets the stale Mock, whose side_effect is the granite_speech5_ctc ValueError from TestUnsupportedArchitectureError. Our own guard converts that into:

ImportError: transformers 5.16.0 cannot load the ASR model 'ibm-granite/granite-speech-5.0-470m-turboctc': its architecture requires transformers>=5.16

which is a false diagnosis — pods serving on transformers 5.16 transcribe correctly. This is the most damaging part: it misattributes unrelated failures to a version problem that does not exist.

Impact: 4 failures + 4 errors on the TurboCTC branch (tests/integration/test_asr_ctc_default_gpu.py, test_audio_serving_smoke[granite-4.0-micro-*]). Latent on main, which has no real CTC load to poison.

Fix: patch transformers.pipeline before transformers.pipelines.pipeline, or force the lazy resolve before patching. Add a regression test asserting the attribute is restored after the context manager exits.