[d-v-b]

🤖 AI text below 🤖

Every public object now has an API reference entry with a docstring, and the struct data type has its reference page.

A new test fails when a public object is undocumented or a public module has no page.

What changed

• Every public module, __all__ name, and member of a public class has a docstring: TypedDict fields, model attributes and methods, the *Update types, the v2 codec parameter and data type configurations, the definition package, the problem fields, the top-level module and __version__. Fields are documented by a string under the field, the style the package already used; the numpy Attributes sections in StructField and the numpy.datetime64/numpy.timedelta64 configurations became per-field docstrings. Spec rules are cited with permalinks pinned to the commits the package already cites; v2 codec parameters (which have no spec) were checked against numcodecs and each module's canonical form.
• zarr_metadata.v3.data_type.struct was on no page; it is now on the data type page, and the API index renders __version__.
• tests/test_api_docs.py reads the package through griffe, as mkdocstrings does, and asserts that every public module has a ::: page, every public object has a docstring, and the on-disk v2 .zarray/.zgroup types carry the same field docstrings as their merged counterparts. griffelib joins the test dependency group.

Where the docs and the code disagree (documented, not changed)

• The package reports must_understand: false on a codec as invalid. The spec forbids it only for data types, chunk grids and chunk key encodings (core/index.rst#L1580-L1581). The codec docstrings state both.
• ZarrV2LzmaParameters says its parameters are "as the lzma module accepts them", but its types admit values lzma refuses when compressing: format: 0, a preset above 9, a preset with filters, a check other than none under FORMAT_ALONE, FORMAT_RAW without filters. The field docstrings describe what lzma accepts.
• The v2 delta and fixedscaleoffset parameters admit dtype: "|O", which numcodecs refuses; bitround admits a keepbits above the mantissa width, which numcodecs refuses when encoding.

Checks

Ran locally: ruff, pyright 1.1.414, the test suite on Python 3.11 and 3.14, and mkdocs build --strict.

Merge notes

The branch merges main after zarr-developers#4490 restructured the package. Conflicts were resolved in favor of main's text inside each conflicting hunk; the docstrings those hunks dropped are restored in the commit after the merge.

🤖 Generated with Claude Code

[d-v-b]

🤖 AI text below 🤖

Superseded by zarr-developers#4590, which merged this branch upstream. The diff against main is now empty.