v0.4.2

Major Features

• Cancel now stops processing. process_fn runs in a reusable worker process, so clicking
  Cancel in HARP terminates the job instead of leaving it to finish server-side.
  build_endpoint also takes timeout_s (default 900). gr.Progress, gr.Info, and
  gr.Warning are forwarded back out of the worker.
• Model tagging. Model cards declare Category, Subcategory, SampleRate, and
  Channels tags from a taxonomy shared with HARP, and input and output tags are inferred
  from the Gradio components. HARP categorizes and searches by these, and a Space can list
  them in its README to be categorized before it is loaded.
• Usable error messages. Gradio 6.13 or newer is required. Earlier versions discard the
  error payload on the endpoint HARP uses, leaving a failed job with no message at all.
• More components. Number boxes, multi-select dropdowns, and generic file inputs and
  outputs.
• Test suite. pytest tests covers the model card, tags, components, labels, media, and
  the worker process.

Minor Improvements

• Expanded README: deployment guidelines, Docker Spaces for dual-environment models,
  binary file storage on the Hub, and self-hosted endpoints over an SSH tunnel.

v0.4.1

Fixed issue with daemon workers for ZeroGPU compatibility.

v0.4.2

Added warm start to save ZeroGPU usage and intuitive output labels.

Breaking Changes

• The Gradio code must sit behind if __name__ == "__main__":, with process_fn defined
  above it. The worker imports the app module to reach process_fn, which has to be
  top-level and picklable.
• ModelCard is keyword-only, and author now precedes description.
• SpectrogramLabel has been removed.