feat(py): PyO3 bindings (fleischwolf-py) — docling-shaped drop-in, local-only

[artiz]

Summary

The strangler-fig PyO3 bindings from the roadmap — experimental, not published anywhere (the PyPI name is tentative), built locally with maturin:

# was:  from docling.document_converter import DocumentConverter
from fleischwolf import DocumentConverter

result = DocumentConverter().convert("document.pdf")
print(result.document.export_to_markdown())
data = result.document.export_to_dict()          # docling JSON wire format, schema 1.10.0

• API surface mirrors docling: DocumentConverter(strict, fetch_images, artifacts_path), .convert(path) (str/os.PathLike, GIL released for the conversion), .convert_bytes(name, data) (the DocumentStream counterpart), ConversionResult.status/.document, DoclingDocument.export_to_markdown/export_to_dict/export_to_json/save_as_*.
• Models download in pure Python, docling-style: fleischwolf.download_models() fetches the pipeline assets (incl. pdfium and the INT8 variants) from the models-v1 GitHub release into ~/.cache/fleischwolf ($FLEISCHWOLF_CACHE_DIR overrides) with stdlib urllib; ensure_env() — called by the constructor — points the native pipeline at the cache via the standard DOCLING_*/PDFIUM_* env vars, preferring INT8 like the Rust default and never overriding explicitly-set variables. Declarative formats need no models at all.
• Deliberately isolated: standalone crate (own empty [workspace], publish = false) — outside the repo workspace, untouched by CI's cargo test/clippy and by ci_publish.sh. Renaming the distribution later costs nothing.
• Crate README documents the full local try-it flow (venv → maturin develop --release → download_models() → convert); MIGRATION §6 moves PyO3 from "planned" to "experimental in-repo".

Testing (all in-container, per the README steps)

• maturin develop --release builds and installs into a fresh venv.
• download_models() pulled 623 MB from the release (required + optional INT8/.data assets, 404-tolerant optional path exercised earlier against a tag without int8).
• Declarative Markdown conversion, export_to_dict() (schema 1.10.0), bytes API and pathlib.Path inputs pass without any models configured.
• A PDF converted through the full ML pipeline from a foreign CWD with a clean environment is byte-identical to the CLI's output on the same file (INT8 stack auto-selected from the cache).

🤖 Generated with Claude Code

https://claude.ai/code/session_017fKFbatf55BK5j8vr3dZjj

---

Generated by Claude Code