fix(whisperx): keep the transcript when diarization fails, report real errors

[walcz-de]

Description

AudioTranscription in the whisperx backend caught every exception and returned an empty TranscriptResult. A failure in the diarization step therefore discarded a transcript that was already finished. Since /v1/audio/transcriptions treats a missing diarize field as true, every request reaches that step; when the HF token has not accepted the terms of the gated pyannote/speaker-diarization-community-1 pipeline, the download fails with 403 and every transcription came back as an empty text with HTTP 200 (backend log: language detected, alignment done, then GatedRepoError, result empty).

Diarization now degrades: if it fails, the transcript is returned without speaker labels and the reason is logged. Any other failure aborts the call with INTERNAL instead of returning an empty, successful-looking result.

test_transcript_utils.py covers the new diarize_or_keep helper: a refused diarization keeps the transcript and logs the error; a successful one is returned.

Notes for Reviewers

A related inconsistency I did not change: the endpoint enables diarization unless diarize=false is sent (relied on by parakeet-cpp since #12335), while whisperx rejects diarization without HF_TOKEN (#8744de44d). So a whisperx install without a token fails every plain transcription with FailedPrecondition. Making whisperx skip diarization when it was not explicitly requested would need the request to carry whether the field was set; I left that decision to you.

Signed commits

• Yes, I signed my commits.
• Documentation updated (docs/content/) for user-facing changes, or not applicable

🤖 Generated with Claude Code

[localai-org-maint-bot]

@mudler code review looks good at d07bf5f80 against master 53cd6c571, subject to the required CI and DCO checks.

The diarization fallback preserves the available transcript, while transcription failures now return a gRPC error. The missing-token precondition remains separate. I added documentation for the fallback and error behavior.

Validation: Static review and git diff --check passed. Python tests could not run because the installed Python executable requires a runtime loader unavailable in this environment.

The documentation commit is pushed to this PR branch. I left its human DCO sign-off to the contributor. No merge performed.

[walcz-de]

Thanks for the review and for adding the documentation. That doc commit had no DCO sign-off, so the DCO check stayed red. I folded equivalent documentation (reworded and checked against the code) into the signed-off fix commit and force-pushed the branch. The code and tests are unchanged from the reviewed state. Sorry for leaving the docs out of the original submission.