HuggingFace inference operator: task dropdown offers tasks with no codegen, which silently fall back to text-gen and emit raw JSON

[PG1204]

What happened?

The HuggingFace operator's task selector lists pipeline tasks the operator can't actually run. STATIC_TASK_OPTIONS in frontend/src/app/workspace/component/hugging-face/hugging-face.component.ts (and the dynamic list fetched from HF's /api/tasks) include tasks that have no corresponding TaskCodegen, e.g. text-classification, token-classification, translation, summarization, feature-extraction, fill-mask.

The operator only registers codegens for a fixed set of tasks (TextGenCodegen, ImageTaskCodegen, AudioTaskCodegen, MediaGenCodegen, QaRankingCodegen, see registeredCodegens in HuggingFaceInferenceOpDesc). Any task outside that set falls through to TextGenCodegen, whose payload hits the generic {"inputs": prompt_value} branch and whose parser only handles text-generation. The result: the row's output cell gets a raw json.dumps(body) of whatever the endpoint returned, with no indication the task isn't supported. The unrecognized-task fallback is intentional as a safety net (keeps codegen total), but the UI actively steering users into it is not.

Proposed fix: prune the task selector to the tasks that actually have a codegen, and filter both STATIC_TASK_OPTIONS and the dynamic /api/tasks result against a supported-task set. "Only show what works" is simpler and clearer than showing unsupported tasks with a warning label. The supported set currently lives implicitly in the Scala codegen registrations, so the frontend needs its own copy of that list; the PR should keep the two in sync (and note it) so a newly added codegen surfaces in the UI. Scope is frontend-only (hugging-face.component.ts + its spec).

Supported tasks today (the allowed set): text-generation; image-classification, object-detection, image-segmentation, image-to-text, visual-question-answering, document-question-answering, zero-shot-image-classification, image-text-to-text, image-to-image; automatic-speech-recognition, audio-classification, text-to-speech; text-to-image, text-to-video; question-answering, table-question-answering, zero-shot-classification, sentence-similarity, text-ranking.

How to reproduce?

1. Add a HuggingFace inference operator and open the task dropdown.
2. Select an unsupported task, e.g. summarization or translation, pick any model, and
   run the workflow.
3. The result column contains a raw JSON dump of the endpoint response instead of a
   task-appropriate result, and nothing warned that the task isn't supported.

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