Audit and improve Hugging Face assets and documentation

[gonzalobenegas]

Outcome

Make published Song Lab assets discoverable, accurately documented, and compatible with the installed package.

Acceptance criteria

• Inventory relevant models, datasets, tokenizers, score resources, and collections; classification does not imply support for every historical asset.
• Record immutable approved revisions and required files in a machine-readable manifest.
• Repair supported model/dataset cards, including install/registration snippets, intended use, inputs, outputs, limitations, citations, licenses, and provenance.
• Keep canonical inference/model code in this repository/PyPI package; retain remote-code fallbacks only when deliberately audited and documented.
• Run one dated compatibility audit covering reachability, config/architecture/tokenizer/assets, portable paths, safetensors, metadata, and representative numerical inference.
• Do not add recurring weekly Hub CI. Rerun the audit only for deliberate compatibility/model-asset changes.
• Prepare card/collection updates for final review; do not mutate Hub repositories until approved.