HuggingFace inference operator: uploaded audio is stored as a server temp path that the Python worker refuses to read, so audio upload never works

[PG1204]

What happened?

The audio-upload flow is broken end to end:

• The frontend posts the file to /api/huggingface/upload-audio, which streams it to ${java.io.tmpdir}/texera-hf-audio and returns the absolute server path (HuggingFaceModelResource.scala ~L171-177). The upload component stores that path in the operator's audioInput property (hugging-face-audio-upload.component.ts ~L118).
• The generated Python's _read_audio_input (PythonCodegenBase.scala ~L797-811) deliberately accepts only data: and https URLs and raises for anything else (filesystem reads were removed as LFI hardening).

Nothing converts the path in between, so every ASR / audio-classification run with an uploaded file fails per-row with "Audio task configuration error: Could not read audio input (ValueError: Unsupported audio input ...)".

The path design is unsound even aside from that: the Python worker need not share a filesystem with the web server in distributed deployments, and the temp file is swept after 1 hour (AUDIO_TEMP_TTL_MS) and on JVM exit (deleteOnExit), so saved workflows hold dangling references. The leftover file-extension map in _get_audio_content_type suggests filesystem support was removed on the Python side without updating the upload flow. Image upload works because that component inlines a data:image/... URL instead.

Fix requires a design decision:

• Option A: match the image component: read the file in the browser and store a data:audio/... URL in audioInput. Simple, drops the temp-file endpoints entirely; cost is workflow-JSON bloat (uploads are capped at 25 MiB today).
• Option B: store uploads somewhere the worker can reach (e.g. dataset storage) and save a URL that the existing hardened https fetch path can consume.

Either way, /upload-audio, /audio-preview, and the preview component must be updated to match, plus spec coverage for the chosen format.

How to reproduce?

1. Add the operator with task automatic-speech-recognition and upload an audio file in the property panel (the saved value is a server path like /tmp/texera-hf-audio/hf-audio-XXXX.mp3).
2. Run the workflow: every row's result column shows "Could not read audio input (ValueError: Unsupported audio input ...)".

Version/Branch

1.3.0-incubating-SNAPSHOT (main)

Commit Hash (Optional)

No response

What browsers are you seeing the problem on?

No response

Relevant log output

[PG1204]

/take