fix(import): skip existing content-addressed .dir objects (0.19.1)

[johnyaku]

Closes #175.

The bug was one function

Your diagnosis was exactly right, and the codebase already had the pattern to copy — the other four .dir writers guard correctly and only create_dir_file didn't:

Site	Before
import_data.py:235, :366, :430	if not dest_file.exists(): … write ✅
update.py:367	same, under the workspace lock ✅
import_data.py:594	unconditional write ← the bug

Cache objects are content-addressed, so an object that already exists is the one being written. It now skips.

Two halves, independently sufficient

1. Skip when the object exists — the proposed fix.
2. Treat a failed write as success if the object exists afterwards — for the race where concurrent --csv imports populate one registry and another writer lands between the check and the write.

I found out these are independent by accident, and it's worth recording. My first mutation test removed only the pre-check, and the reproduction test still passed — because the except OSError recovery handles the permission case on its own. Reverting both halves, it fails properly. So the belt and the braces each do the job; keeping both covers the concurrent case the --csv bulk form makes likely.

One addition beyond the issue

When skipping, the existing object's size is compared against what would have been written, and a mismatch warns.

A truncated .dir keeps its filename, so it still looks like the correct object while decoding to a short file list — surfacing much later as a fetch that silently misses files. Two syscalls, and it warns rather than fails, so it cannot break an import that currently works.

This deliberately diverges from the other four sites, which skip silently. Happy to drop it if you'd rather they stayed consistent.

Verification

8 new tests. 4 fail against the unfixed original, including test_read_only_object_no_longer_fails, which reproduces #175 with real chmod 0444 on the object and 0555 on its directory rather than a mocked PermissionError.

I also confirmed independently that a 0444 file genuinely raises PermissionError for its owner on this filesystem, rather than assuming owner-bypass semantics — that assumption is the sort that quietly makes a reproduction test vacuous.

Full suite: 1997 passed, 1 skipped.

Also

• docs/import.md gains a troubleshooting entry. "Worked for my sections, failed for my colleague's" reads as intermittent rather than as a permissions problem, and that's worth naming.
• Not done here: retiring the import_data.sh dvc import fallback the issue mentions — that lives in swarbricklab/images, not this repo. Worth a follow-up there once this ships.

🤖 Generated with Claude Code