fix(polyglot): bound recursive archive extraction and unify error handling

[feiiiiii5]

Root cause

find_file_properties_recursively exists to inspect untrusted archives, but its zip and tar paths extracted members to disk with no size limit and no error handling: zipped_file.read(fname) streams the full decompressed output, so a compressed-expansion bomb (tiny archive, huge inflated member) exhausts storage during scanning. The 7z path already handled both concerns (BytesIOFactory(limit=...) + except (OSError, ArchiveError) graceful degradation) — the zip and tar paths matched neither.

A second defect in the same loop: entries written by zipfile.writestr carry permission bits without file-type bits (external_attr = 0o600 << 16, so mode == 0 is false and stat.S_ISREG(mode) is false) — those members were silently never scanned.

Also removes a stray debug print("check_if_legacy_format") that fired on every tar-file format check through identify_pytorch_file_format → PyTorchModelWrapper.validate_file_format.

Fix

• Zip/tar member extraction streams in chunks up to MAX_MEMBER_EXTRACTION_SIZE, raising ResourceExhaustionError past the cap — declared archive sizes lie, so the bound is enforced while streaming rather than read from metadata; this aligns the polyglot scanner with the expansion-attack protection the analysis path already has
• Corrupt members degrade gracefully (BadZipFile/TarError/OSError → member skipped), matching the 7z path
• Zip member filter keys off the file-type bits (S_IFMT), so permission-only entries (writestr style) are scanned
• Stray debug print removed

Validation

• pytest test/test_polyglot.py | sample size: N=21 | key metrics: 4 new tests fail pre-fix (extraction uncapped on a 4 MiB-of-zeros member with a 1 KiB cap; stray print present; writestr members skipped) → post-fix 21 passed | result: pass
• Full suite pytest test/ → 153 passed
• ruff@0.16.0 check + format --check: clean

Risk / Compatibility

• Well-formed archives behave identically (the cap default is 1 GiB per member); hostile archives now fail loudly with ResourceExhaustionError instead of filling the disk, and corrupt members degrade instead of crashing — both align zip/tar with the 7z path's existing contract.

[feiiiiii5]

Hi @thomas-chauchefoin-tob, the #325 branch is still focused on bounded archive extraction and consistent error handling, with no changes since the last update. Is there a remaining concern you’d like addressed before review, or should I leave the current scope unchanged?

[feiiiiii5]

Second bump on this one, sorry — the only review activity so far has been my own first ping on 24 September.

The scope has not changed since then, so rather than repeat it: the two things I would most appreciate a maintainer ruling on are

1. the extraction limit — the 7z path already used BytesIOFactory(limit=...), and I matched that shape for zip and tar. If the project wants a different ceiling, or wants it configurable rather than fixed, that is a one-line change on my side and I would rather match your convention than guess at a number.
2. the permission-bits hunk. Entries written by zipfile.writestr with external_attr = 0o600 << 16 have no file-type bits, so stat.S_ISREG(mode) is false and those members were silently never scanned. I fixed the test rather than the reader. If you would rather the reader treat "no file-type bits but readable" as a regular file, say so — that is arguably the more correct fix and it is not much larger.

If neither is a concern, I would also rather you close it than leave it hanging — happy to hand the bounded-extraction fix over as-is if the permission part is contentious, or vice versa.

Either way, thanks for the time.