Add Baidu Unlimited-OCR engine adapter

[Mihailorama]

Summary

Evaluates and adds Baidu Unlimited-OCR (the project from the uploaded screenshot) as a new docfold engine.

Unlimited-OCR is an open-weight (MIT) document-parsing Vision-Language Model released June 2026. It builds on DeepSeek-OCR and replaces the decoder attention with Reference Sliding Window Attention (R-SWA), giving a constant KV cache so it can transcribe dozens of pages in a single 32K-context forward pass — its "one-shot long-horizon parsing" headline. It's free, locally runnable, and a good fit alongside docfold's existing local VLM engines (Chandra, Surya).

The new UnlimitedOCREngine wraps the upstream HuggingFace model (baidu/Unlimited-OCR, loaded via trust_remote_code), mirroring the existing ChandraEngine/SuryaEngine adapters:

• Lazy model load (never at import / construction / is_available() / registration time); blocking inference runs in run_in_executor.
• gundam (default) and base modes mapped to upstream base_size/image_size/crop_mode.
• Accepts images directly; renders PDF pages to images via PyMuPDF.
• Emits Markdown / HTML / JSON / text via the unified EngineResult.

Built per the project's mandatory TDD workflow: proposal → failing tests → implementation → verification.

Changes

• New: src/docfold/engines/unlimited_ocr_engine.py
• Router: added unlimited_ocr to PDF + image priority lists and default fallback (src/docfold/engines/router.py)
• CLI: registered in _build_router() (src/docfold/cli.py)
• Packaging: new unlimited-ocr optional-dependency group, added to [all] (pyproject.toml)
• Tests: TestUnlimitedOCREngine (10 cases) + interface parametrize entry (tests/engines/test_adapters.py)
• Docs: proposal (docs/tasks/UNLIMITED_OCR_ENGINE.md), README engine tables, benchmarks profile, CHANGELOG

Testing

• pytest tests/ → 329 passed, 3 skipped
• ruff check and mypy on new/changed files → clean
• Smoke-tested router registration: engine registers, advertises pdf+image extensions and table/heading/reading-order capabilities (is_available() is False until torch/transformers are installed)

Notes

• The model requires a CUDA GPU and heavy deps (torch, transformers, einops); these only load when the engine actually runs, so the adapter imports cleanly without them.
• First cut processes PDFs page-by-page for determinism/testability. Native single-pass infer_multi long-horizon batching is noted as a future enhancement in the proposal.
• Repo convention (docs/conventions/golden-rules.md) says "never push"; this session's operator instructions explicitly require pushing the feature branch and opening a draft PR for review, which preserves that intent (no direct-to-main, no release).

🤖 Generated with Claude Code

https://claude.ai/code/session_01P23UQav6r1QsPGKzogma26

---

Generated by Claude Code