Fix v2 structured field selection

[puneetdixit200]

Fixes #3983

This fixes field selection for Zarr v2 arrays with structured dtypes. The read path now decodes chunks into a full-dtype buffer when fields are requested, then copies the selected field view into the caller-facing output buffer. That avoids trying to assign full structured records directly into a single-field or subset-field output dtype.

I added a regression test covering direct string field indexing, basic selection with a single field, and multi-field selection for v2 structured arrays.

AI assistance: OpenAI GPT-5 was used during implementation; I reviewed the change and ran the checks below.

TODO:

• Add unit tests and/or doctests in docstrings
• Add docstrings and API docs for any new/modified user-facing classes and functions
  • Not applicable: no user-facing API was added or changed.
• New/modified features documented in docs/user-guide/*.md
  • Not applicable: this is a bug fix.
• Changes documented as a new file in changes/
• GitHub Actions have all passed
  • Pending after PR creation.
• Test coverage is 100% (Codecov passes)
  • Pending after PR creation.

Local checks:

• .venv/bin/python -m pytest tests/test_v2.py -q
• .venv/bin/python -m pytest tests/test_v2.py::test_structured_dtype_field_selection tests/test_indexing.py::test_get_selections_with_fields -q
• .venv/bin/python -m ruff check src/zarr/core/array.py tests/test_v2.py
• .venv/bin/python -m ruff format --check src/zarr/core/array.py tests/test_v2.py
• git diff --check

[puneetdixit200]

Pushed a follow-up for the min-deps CI failure. The failure was caused by package version discovery selecting the auxiliary zarr_metadata-v0.2.0 tag, which made the in-tree package report 0.2.1.dev...; numcodecs.zarr3 then rejected it as < 3.0.0.

Additional local checks:

$ UV_CACHE_DIR=$PWD/.uv-cache uv run --with hatch --with hatch-vcs hatch version
3.2.2.dev30+ge9e83506.d20260521

$ UV_CACHE_DIR=$PWD/.uv-cache uv run --with 'numcodecs==0.14.1' python - <<'PY'
import zarr
print(zarr.__version__)
import numcodecs.zarr3
print('numcodecs.zarr3 import ok')
PY
3.2.2.dev30+ge9e83506.d20260521
numcodecs.zarr3 import ok

$ UV_CACHE_DIR=$PWD/.uv-cache uv run --with 'numcodecs==0.14.1' pytest tests/test_codecs/test_numcodecs.py::test_get_codec_class -q
.....................                                                    [100%]
21 passed in 0.05s

$ UV_CACHE_DIR=$PWD/.uv-cache uv run pytest tests/test_v2.py::test_structured_dtype_field_selection tests/test_indexing.py::test_get_selections_with_fields -q
.x                                                                       [100%]
1 passed, 1 xfailed in 0.27s

$ UV_CACHE_DIR=$PWD/.uv-cache uv run ruff check pyproject.toml src/zarr/core/array.py tests/test_v2.py && git diff --check
All checks passed!

[d-v-b]

hi @puneetdixit200 before you get too deep in to this PR, it might make sense to resolve the points I raised in #3983

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.54%. Comparing base (c0e2afa) to head (05be8b3).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3996      +/-   ##
==========================================
+ Coverage   93.49%   93.54%   +0.04%     
==========================================
  Files          88       88              
  Lines       11873    11877       +4     
==========================================
+ Hits        11101    11110       +9     
+ Misses        772      767       -5     

Files with missing lines	Coverage Δ
src/zarr/core/array.py	97.79% <100.00%> (+<0.01%)	⬆️

... and 1 file with indirect coverage changes

🚀 New features to boost your workflow:

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

[puneetdixit200]

Thanks, I read through #3983. I agree the API question around string __getitem__ is the important part here.

The code change is currently in the shared fields selection path, so it also fixes explicit calls like get_basic_selection(..., fields="age"). The test includes za["name"] because that is the v2 regression reported in the issue, but I can narrow the PR if the preferred direction is to support only the explicit fields= selection APIs and leave string __getitem__ out for array API compatibility.

I will hold off on further code changes until there is a clearer direction on that API choice.