Native-first Python binding with OCR and an OCRmyPDF plugin

[Frooodle]

Summary

Python bindings to JPDFium's native core, with no JVM. This is a thin nanobind layer over the safe jpdf_* API from Stirling-Tools/JPDFium feat/native-safe-domain, which CI builds from source. One stage per commit:

1. Lazy-loading binding. Documents, pages, rendering, bitmaps and zero-copy views. The GIL is released around native calls.
2. Rasterization and rectification matching OCRmyPDF's rasterizer.
3. Text layers and page text, accepted by OCRmyPDF's own sandwich graft.
4. OcrEngine: PP-OCRv6 via JPDFium's OCR module, matching the official PaddleOCR pipeline.
5. Right-to-left, mixed-direction and rotated text layers, qualified across four text extractors.
6. The pyjpdfium command and the OCRmyPDF plugin.
   • Commands: info, render, ocr, fonts, models, doctor.
   • The plugin, ocrmypdf --plugin pyjpdfium.ocrmypdf, needs neither Tesseract nor Ghostscript. It refuses modes it has not been checked in, PDF/A among them.
7. Self-contained wheels (tools/build_wheel.py).
   • They are built from JPDFium's own bundler, carry an identity manifest and every bundled license, and are checked on a clean offline Ubuntu.
8. OcrEngine(max_concurrent=...), plus benchmarks against Tesseract (docs/PERFORMANCE.md).

The native core is staged with its own sonames (tools/stage_natives.py). Without that, OCRmyPDF's pypdfium2 would capture our PDFium dependency in the same process.

Testing

All offline in containers:

• Unit and native suites: 110 tests, with both OCR profiles.
• Parity:
  • OCRmyPDF's rasterizer and graft: 57 raster cases and 4 text layer checks.
  • Multilingual text layers: 7 checks.
  • PaddleOCR: 8 of 8 cases for each profile.
• OCRmyPDF end to end:
  • 21 runs, without Tesseract or Ghostscript installed;
  • scanned and rotated pages; skip, force and redo modes; forked workers;
  • exact-image masking, and every refusal.
• Clean ubuntu:24.04: the wheel and the plugin run with no network, and every native library loads from the wheel.

Notes

• The new CI steps have not run in GitHub Actions yet: OCRmyPDF end to end, the clean-machine wheel check, and the second profile's parity.
• On clean English scans, Tesseract is faster and smaller at the same accuracy (docs/PERFORMANCE.md).
• The plugin's language support is English only, the language checked end to end.