fix: keep element order when writing F-contiguous chunks with vlen codecs

[oldrobotdev]

Summary

Fixes #3558. Writing an F-contiguous object-dtype array to a chunked string or bytes array scrambles the stored data whenever a chunk is covered exactly by the selection: the write path passes the chunk view to numcodecs without a copy, numcodecs flattens it with order='A' (column-major for F-contiguous input), and decode reshapes in C order. The wrappers in codecs/vlen_utf8.py now pass np.ascontiguousarray(...) to the codec, a no-op for C-contiguous chunks and a copy of the object-pointer array otherwise. Regression tests cover vlen-utf8 and vlen-bytes at both the exact-cover and partial-chunk shapes; the exact-cover cases fail on current main.

For reviewers

The one design question is the fix layer: this could live in numcodecs instead (change the flatten to C order), but that touches the v2 out= decode path where the old behavior was self-consistent for F-ordered arrays, so I kept the change on the zarr side where decode already commits to C order. The repro and codec-level probe are in #3558.

Author attestation

• I am a human, these are my changes, and I have reviewed and understood every change and can explain why each is correct.

TODO

• Add unit tests and/or doctests in docstrings
• Add docstrings and API docs for any new/modified user-facing classes and functions (no user-facing API change)
• New/modified features documented in docs/user-guide/*.md (not applicable)
• Changes documented as a new file in changes/
• GitHub Actions have all passed
• Test coverage is 100% (Codecov passes)

[codecov]

Codecov Report

❌ Patch coverage is 50.00000% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 93.58%. Comparing base (4b68c2c) to head (c0be9f4).

Files with missing lines	Patch %	Lines
src/zarr/codecs/numcodecs/_codecs.py	50.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #4116      +/-   ##
==========================================
+ Coverage   93.50%   93.58%   +0.08%     
==========================================
  Files          90       90              
  Lines       11981    11981              
==========================================
+ Hits        11203    11213      +10     
+ Misses        778      768      -10     

Files with missing lines	Coverage Δ
src/zarr/codecs/vlen_utf8.py	92.42% <ø> (+15.15%)	⬆️
src/zarr/codecs/numcodecs/_codecs.py	96.38% <50.00%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[d-v-b]

thanks for this fix! Could you check if the other numcodecs codecs (delta, fixedscaleoffset, etc) are affected by the same bug?

[oldrobotdev]

Checked, and yes: Delta, FixedScaleOffset and PackBits all round-trip transposed data for an F-contiguous chunk that exactly covers the selection, both when calling the numcodecs codec directly and through a zarr array with the filter configured. The path is the generic adapter in codecs/numcodecs/_codecs.py, which hands chunk_data.as_ndarray_like() to numcodecs unchanged while decode reshapes in C order, so every array-array filter goes through the same exposed path.

I pushed the same ascontiguousarray treatment for the array-array and array-bytes adapter encode paths, with regression tests for the three filters above and the changelog entry updated to match. One probe gotcha worth mentioning for review: a test pattern that is symmetric under transpose passes even on broken code (arange % 3 on a 16-wide array is symmetric, since 16 ≡ 1 mod 3), so the tests use arange and a lower-triangular bool mask.

I'll open the numcodecs issue about the underlying order='A' behavior next, since direct numcodecs users stay exposed either way.