Unlimited-OCR — batch parse a folder of PDFs and images

Process an entire folder of images and PDFs without losing work to interruption, using the same proven baidu/Unlimited-OCR model that powers unlimited-ocr-parse-document. Results are written incrementally to a manifest, so a long run that gets interrupted can be resumed without reprocessing.

Self-Evolving Skill: every number and guarantee below was measured on this hardware, not copied from upstream documentation. If a guarantee stops holding, fix this file and references immediately rather than deferring. Re-verify before changing a documented fact.

The guarantees

1. ONE IMAGE PER FORWARD PASS, ALWAYS — per-image processing is deliberate and measured. Single-pass multi-image on MLX recovered 1/3, 4/5, and 0/10 pages; per-image recovered 100 % every time. See ../../references/PITFALLS.md § 3.
2. CHECKPOINT AND RESUME — every result is written to disk BEFORE the manifest is updated. If the process dies, the incomplete line is never recorded. Re-run with the same output folder to skip everything already done.
3. CHARTS COME BACK EMPTY — this is correct behaviour for a layout parser and a fatal surprise if you expected a vision model. See the next section.
4. EXIT CODES: The batch script returns a summary exit code: 0 if all files succeeded without repetition, 1 if some files had repetition suspected (check before trusting), 2 if any files failed with hard errors. Per-file status is always recorded in the manifest.
5. MEMORY-BOUNDED — no result accumulation. Each file runs in its own process (load, infer, write, exit), so a 1,000-file folder never threatens machine stability.

What it will NOT do