fix(zarr-indexing): reject overflowing integer selectors

[d-v-b]

🤖 AI text below 🤖

Summary

Direct IndexTransform.oindex and .vindex selections cast integer arrays to np.intp before checking whether their values fit. On a domain containing negative coordinates, np.array([2**64 - 1], dtype="uint64") therefore silently selects coordinate -1.

Use the existing checked_affine(0, 1, ...) conversion for direct integer-array and list selectors. Coordinates outside the representable range now raise OverflowError before narrowing; valid coordinates keep their meaning. The positional LazyArray boundary already checks unsigned bounds and is unaffected.

For reviewers

Regression tests cover both direct indexing modes, the signed-range boundary and uint64 maximum, NumPy-array and list inputs, and read-only valid selectors with repeated/unsorted coordinates. Four NumPy-array cases silently succeeded before this fix; list inputs already raised an overflow error and now use the same diagnostic.

Validation: the zarr-indexing source test suite passes with 1,333 passed and 4 skipped (optional TensorStore/Dask unavailable), using the Python 3.12 minimal Hatch environment. Temporary distribution-version metadata was supplied for this source-only checkout, including its subprocess example; packaging was not tested by that run. Ruff lint and formatting pass.

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 regression tests
• Add a subpackage changelog entry
• Human author reviews the changes and edits the description
• GitHub Actions have all passed