test: give the sharded coordinate-selection write path real weight in test_oindex

[d-v-b]

🤖 AI text below 🤖

Summary

#4284 / #4316 fixed orthogonal writes on sharded arrays where the sharding codec re-derives a CoordinateIndexer and must ravel the value, but the property test meant to cover it, test_oindex, reached that path in 2 of 300 examples under the derandomized ci Hypothesis profile — a fixed outcome, so a green run said almost nothing about the fix. The test draws one_of(simple_arrays, rectilinear_arrays), rectilinear arrays cannot shard, and only ~8.6% of simple_arrays draws are sharded.

This adds zarr.testing.strategies.sharded_arrays, a strategy that always yields a sharded Zarr format 3 array (chunk shape, shard shape, subchunk write order and inner codec chain are drawn; the ShardingCodec construction is factored into a private helper shared with arrays() with the same draw order, so existing derandomized sequences are unchanged), and draws it as a third arm in test_oindex and test_vindex. The value-reshape path is now exercised in ~50 of 300 ci examples (13–27% of writes across random seeds) with no measurable change in suite time, and removing the guard in sharding.py makes test_oindex fail (shrunk example: a (1, 3, 1) array with shards (1, 3, 1), selection (array([0]), slice(1, 3), 0), value shape (1, 2)). A Hypothesis event() at the write site makes the reach visible under --hypothesis-show-statistics.

Separately, the docstring of test_set_selection_rejects_value_with_wrong_rank in tests/test_indexing.py claimed storage layout never changes which writes are rejected; it does — a (2, 2, 1) value on oindex[[3, 1], [0, 2]] is accepted by a chunked array (each chunk's piece is (1, 1, 1), which numpy broadcasts) and rejected by a sharded one — so the docstring now states exactly what the test pins.

Found during the pre-release review for #4256.

For reviewers

The design choice worth a look is a dedicated strategy rather than a sharded= knob on arrays(): Zarr format 2 arrays cannot shard, so a knob would have to silently override zarr_formats. Note that the fixed path is reached by a bare-int axis plus one array axis too (OrthogonalIndexer routes through ix_ when n_array_dims > 1 or drop_axes), not only by two or more array axes; the min_dims=2 on the test_oindex arm follows from the guard needing a ≥2-D value. derandomize=True re-rolls the outcome on every edit to the test body, so the ci reach count is only stable while the test is.

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 (sharded_arrays is picked up by the existing zarr.testing.strategies API page)
• New/modified features documented in docs/user-guide/*.md (n/a)
• Changes documented as a new file in changes/
• GitHub Actions have all passed
• Test coverage is 100% (Codecov passes)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 94.32%. Comparing base (542ceba) to head (67e69cd).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #4331      +/-   ##
==========================================
+ Coverage   94.30%   94.32%   +0.01%     
==========================================
  Files          92       92              
  Lines       12915    12941      +26     
==========================================
+ Hits        12180    12206      +26     
  Misses        735      735              

Files with missing lines	Coverage Δ
src/zarr/testing/strategies.py	96.14% <100.00%> (+0.33%)	⬆️

... and 2 files with indirect coverage changes

🚀 New features to boost your workflow:

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

[d-v-b]

consider supporting recursive sharding. A single recursion is likely enough to catch weird issues.