Yue2 "stop after" (generate just score.abc, or semantic)

[engival]

This gives you the ability to generate just a score.abc, or the score + semantic, while skipping the final audio render.

• Use case: score-only and tokens-only runs.
  • Write a score once, then render it in several styles without paying for the NAR and VAE each time.
• Two structural changes:
  • Yue2RunResult::audio became std::optional. Otherwise the CLI writes a header-only WAV.
  • The ABC stage moved verbatim into generate_plan_abc / attach_plan_abc, so the early return can share it.
• Validation on Intel Arc / Vulkan:
  • A request without the option renders a byte-identical WAV to dev.
  • Feeding the exported score back through abc_file reproduces a byte-identical semantic.json.
  • stop_after=abc with cot=off is rejected, and so is stop_after=abc with a supplied score. The error messages are clear.
  • The three YuE2 path cases pass.
• Limits:
  • A no-audio result works through /v1/tasks/run. It does not work through /v1/audio/speech, which requires audio and is left untouched.
  • There is no WebUI control.
  • It was not tested in the WebUI.

[0xShug0]

@engival One issue should be fixed before merging:

stop_after=semantic does not automatically return the generated semantic tokens. With cot=off and no export_semantic=true, it runs semantic generation successfully and returns an entirely empty result: no audio or semantic artifact.

This also conflicts with the model-spec description, which says the semantic setting writes the semantic export. The clean fix is for stop_after=semantic to imply semantic export. At minimum, reject the request when it would produce no output.

[engival]

@0xShug0

Okay, it's done. stop_after now implies the export. The export flag is only required for getting the semantic at the same time as the wav.

[0xShug0]

@engival Thanks! PR merged 🎉