Add browser-local AI workflow repair

[Copilot]

Adds an in-browser prompt experience that uses local AI inference and the WASM compiler to iteratively produce a valid workflow.

• Local inference

  • Lazily loads a pinned Transformers.js runtime and SmolLM2 model in a Web Worker.
  • Keeps inference off the UI thread and validates worker requests.

• Compile-repair loop

  • Applies generated workflow Markdown only after successful compilation.
  • Feeds compiler diagnostics back to the model for up to three attempts.
  • Treats model output as untrusted text and bounds prompt, workflow, diagnostic, and output sizes.

• Loading experience

  • Reports WASM compiler download progress.
  • Shows model download, initialization, generation, and validation progress.
  • Discloses the approximate first-use model download before execution.

• Coverage

  • Verifies inference remains unloaded until requested.
  • Covers compiler-error feedback and successful retry behavior.