fix(dtype): fix nested v2 round trips and detect unsupported field conversions

[d-v-b]

🤖 AI text below 🤖

Fix nested structured dtype round-tripping in V2 metadata, detect field features unsupported by the current native dtype conversion, and warn when that conversion changes layout.

• Validate nested V2 field definitions as lists of fields. The baseline can emit nested metadata that its own dtype parser rejects.
• Reject native NumPy field titles and subarray fields, including nested ones, with a ValueError identifying the field. On the baseline, dtype conversion treats string title aliases as additional fields and subarray fields as raw bytes. Array writes using these inputs can consequently fail with unrelated errors.
• Preserve the baseline's conversion of padded layouts to packed layouts, adding a warning that field values are preserved when writing arrays but offsets and itemsize may change. The integration tests cover aligned layouts, nested padding and explicit offsets in both formats.

Format and implementation boundaries

Subarray rejection is a limitation of the current Zarr-Python dtype implementation, not of V2. The V2 specification explicitly supports [fieldname, datatype, shape]. Nested fields and subarrays were clarified in 2018's #296. A probe using Zarr-Python 2.18.7 successfully writes and reopens a structured subarray field. Restoring that support is outside this change.

For titles, NumPy's descr supports (title, name) pairs. Zarr adopted dtype.descr in April 2016, but the JSON decoder did not reconstruct those pairs as tuples. The clarified V2 grammar describes field names as strings. The reviewed history provides no explicit rationale for excluding titles; this PR makes no claim that they are inherently unrepresentable.

The V3 struct extension separately specifies named, fixed-size fields and packed encoding. Its constraints should not be applied retroactively to V2.

Validation

• 770 dtype, V2 and property tests passed; 5 existing skips.
• New strategies sample built-in dtype classes, bounded parameters and nested layouts. They do not claim exhaustive coverage or support arbitrary third-party dtype constructors. Datetime generation excludes the microsecond alias and normalizes generic-unit scale factors; those normalization issues remain separate.
• The structured-dtype property independently classifies NumPy field descriptors. It requires rejection for the currently unsupported features and successful conversion for the other generated cases, checking packed dtype equality and layout warnings. An injected unrelated ValueError is detected rather than counted as success.
• Compared runtime behavior of Zarr-Python 2.18.7, the PR baseline ba883a556, and this branch. The statements above distinguish those versions deliberately.

Supersedes d-v-b#188.

[d-v-b]

🎵 self-merging when tests pass 🎵

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 94.39%. Comparing base (ba883a5) to head (209a235).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #4340      +/-   ##
==========================================
+ Coverage   94.34%   94.39%   +0.05%     
==========================================
  Files          92       92              
  Lines       12935    13016      +81     
==========================================
+ Hits        12203    12286      +83     
+ Misses        732      730       -2     

Files with missing lines	Coverage Δ
src/zarr/core/dtype/common.py	88.50% <100.00%> (+2.29%)	⬆️
src/zarr/core/dtype/npy/structured.py	93.64% <100.00%> (+1.02%)	⬆️
src/zarr/testing/strategies.py	96.66% <100.00%> (+0.52%)	⬆️

🚀 New features to boost your workflow:

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