feat(python): ship the libmagic database inside the wheel

[andiwand]

🤖 Generated with Claude Code

Wheel builds disabled libmagic on the premise recorded in pyproject.toml:

# Keep wheels self-contained: skip the components whose runtime data cannot be
# bundled into the package.
ODR_WITH_LIBMAGIC = "OFF"

That premise doesn't hold up:

• Not a linking problem. Wheels build with BUILD_SHARED_LIBS=OFF, so libmagic and its zlib/xz/bzip2 dependencies link into pyodr/_core.<abi>.so. Nothing for auditwheel to vendor.
• Not a size problem. magic.mgc is 8.5 MB raw but deflates to ~0.4 MB, and wheels are zip-deflated.
• Not a packaging problem. python/CMakeLists.txt already installs the whole build data directory as pyodr/data, so ODR_BUNDLE_ASSETS=ON is enough to put the database there.

What was actually missing is the runtime fix-up. The compiled-in default from project_info::libmagic_database_path() is the relative string "share/magic.mgc" — meaningless inside a site-packages install — and pyodr/__init__.py only ever repointed the core data path. This gives the database the same three-step resolution already used there (already-configured path → ODR_LIBMAGIC_DATABASE_PATH → bundled copy), keyed on isfile rather than isdir.

What this buys

odr::mimetype() is the only consumer of libmagic (src/odr/internal/magic.cpp:99); the decode pipeline uses odr's own sniffing regardless. So the effect is precise: MIME detection stops bottoming out at the container type.

	before	after
pyodr.mimetype("x.odt")	application/zip	application/vnd.oasis.opendocument.text

Verification

Built a real wheel and installed it into a fresh venv:

magic db: .../site-packages/pyodr/data/magic.mgc
isfile  : True
33 passed

• pyodr/data/magic.mgc present in the wheel: 8,513,392 raw → 443,589 compressed; wheel total 3.29 MB compressed / 18.2 MB installed.
• In-tree build (the build_test.yml path, PYTHONPATH=build/python): 33 passed — pyodr_package already mirrors the data directory, so the database resolves there too.
• With magic.mgc removed from the staged tree: 32 passed, 1 skipped — the new test degrades to a skip rather than a failure, so builds without a database stay green.

Notes

• python.yml drops -o '&:with_libmagic=False' (conan already defaults it on) and adds -o '&:bundle_assets=True', which is what bridges LIBMAGIC_DATABASE_PATH from the dependency runenv into the build.
• pdf2htmlEX and wvWare stay off — their runtime data genuinely is a different story.
• Nothing missing is fatal: libmagic.cpp:35 already falls back to the system database, and magic.cpp:101 then falls back to odr's own sniffing.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 53f65a9744

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".