MuseTalk (Tencent Music Entertainment Lyra Lab) — vendored inference source.

Upstream:  https://github.com/TMElyralab/MuseTalk
Commit:    0a89dec45a0192b824e3cf4daf96c239440c5ed8  (2025-09-26)
Licence:   MIT — see MUSETALK-LICENSE

What is here and why: `musetalk/` (the UNet wrapper, the VAE, face parsing, the
DWPose preprocessing and the vendored OpenAI Whisper audio encoder),
`scripts/inference.py` and `configs/` (the task-list templates). Training, the
realtime demo, the Gradio app and the sample media are not here.

Not cloned at install time — see LATENTSYNC-PINNED for the reasoning. The WEIGHTS
are fetched from the Hub on first use; upstream's download_weights.sh takes the
BiSeNet face-parsing checkpoint off a Google Drive id, which coderai does not use:
it comes from a Hub mirror so the whole set has one provenance.
