Nothing detects drift between polyfile/magic_defs and the file submodule #3479

Summary

polyfile/magic_defs/ is a hand-maintained copy of file/magic/Magdir/ from the file submodule, and nothing verifies that the two agree. A definition can be deleted, truncated, or left behind at an older revision, and the whole test suite stays green.

This matters because the copy is not append-only in either direction:

• Stale definitions accumulate. The update procedure has historically been cp file/magic/Magdir/* polyfile/magic_defs/, which never removes what upstream deleted. Six definitions that libmagic dropped between 2021 and 2024 — alpha, dsf, guile, neko, pc88, rinex — were still shipped and still matching until the libmagic 5.48 update removed them.
• Local patches get reverted. polyfile/magic_defs/c-lang carries the rewrite from Example files that cause an unexpected delay iterating (infinite?) matches  #3411, so it deliberately differs from upstream. A plain copy silently reverts it, reintroducing a matcher that does not terminate on CRLF C++ headers. tests/test_magic.py::test_cpp_class_test_terminates does happen to catch that one, but only because someone wrote a targeted test for it; nothing generalizes.

Reproducer

Delete a bundled definition that the submodule still ships:

rm polyfile/magic_defs/partclone
pytest tests/test_magic.py -q

Output on master (commit eed1f27, with the 5.48 definitions):

21 passed, 91 subtests passed in 5.59s

partclone declares no !:mime, so even the MIME type count is unchanged at 895. The definition is simply gone.

A definition carrying text tests would trip test_text_tests by accident, since that golden set is keyed by file name and line number, but that is a side effect of an unrelated assertion rather than a drift check.

Suggested check

A test in the spirit of tests/test_licensing.py, which already guards an analogous invariant between MANIFEST.in and the Kaitai license allowlist:

1. Every file in file/magic/Magdir/ exists in polyfile/magic_defs/.
2. The only extra entries in polyfile/magic_defs/ are the five PolyFile owns: __init__.py, COPYING, csv, json, polyfile_zip.
3. Every shared file is byte-identical, except for an explicit, documented allowlist of locally patched files. That allowlist currently holds one entry, c-lang, and each entry should name the issue that justifies it.
4. polyfile/magic_defs/COPYING matches file/COPYING.

Point 3 is the valuable one: it turns "someone remembered to re-apply the patch" into something CI enforces, and it makes the set of local patches discoverable instead of folklore. It also fails loudly when upstream rewrites the same lines a patch touches, which is the moment a maintainer needs to think rather than copy.

The test needs the submodule checked out. CI already does submodules: recursive, and test_magic.py::test_file_corpus already asserts file/tests exists, so a skipUnless on file/magic/Magdir would match existing practice for local runs without it.

docs/updating_libmagic_defs.md, added by the libmagic 5.48 update, documents the manual
verification commands. This issue is about making CI run them.

Environment

• PolyFile at commit eed1f27
• libmagic 5.48