Playground: snap hover bboxes to the PDF text layer + show partial extractions

[RATCHAW]

What

Improves the playground's killer interaction — hover a JSON field → highlight its source region — and makes failed runs more useful.

Hover bbox accuracy (snap.ts)

Rescues hover highlights whose model-reported provenance is missing or wrong by locating each field's value in the PDF text layer:

• Builds visual lines from positioned text runs (reading order, gap-aware spacing).
• Matches the value's textual forms, including formatted currency (6000 → $6,000.00), case-insensitively, rejecting matches glued to other word characters.
• Snaps to the occurrence nearest the model bbox, searching across pages to fix null/out-of-range page indices (e.g. 1-based reporting).
• Falls back to the model's own bbox when the text can't be located. Images (no text layer) keep drawing the model bbox directly.

Partial extractions on error

On a MissingRequiredFieldsError, the server now carries the extracted-so-far data/fields/usage on the ApiError, and the client renders it under the error banner instead of showing nothing.

Misc

• Add Gemini 3.5 Flash to the playground model catalog.
• Add the demo GIF to the README.

Tests

New snap.test.ts covers candidate string generation, snapping, cross-page location, and rescue cases. app.test.ts covers the partial-extraction error path.

🤖 Generated with Claude Code