[wpyaoxueros]

Description

Codex on behalf of @wpyaoxueros. Closes #11739.

Contributor confirmation, translated from Chinese: "I have reviewed the changes and found no issues."

CuPy indexers currently fail when they reach pandas-backed coordinates, basic assignment can mix NumPy values with CuPy storage, and scalar truth conversion can trigger an implicit conversion to NumPy. This patch explicitly converts indexers at the pandas boundary, converts assignment values to the target array backend, and uses the underlying array's truth conversion while preserving pandas extension-array handling.

Regression coverage includes vector/multidimensional/argmax indexers, scalar/slice/.loc assignment, scalar CuPy truth values, idxmin/idxmax, and pandas nullable-boolean controls.

Remote validation against base fbc4cf19b919ce23fd0f7d419e9c141f046788a9:

• Same added test cases: 10 failed, 1 passed on unchanged source.
• Patched CuPy/bool selections: 20 passed.
• Related indexing/index/variable modules: 808 passed, 67 skipped, 8 xfailed, 3 xpassed.
• Ruff checks/formatting on the six changed files and diff whitespace checks passed.

These checks ran on the designated Linux GPU workstation. The full repository suite and type checking were not rerun on this commit. The required release-note entry has been added. The contributor has confirmed review of the changes; the remote checks below were executed by Codex.

Documentation CI on 6cbb300 failed in the unchanged doc/internals/time-coding.md timestamp example with pandas 3.1.0. A standalone remote pandas-only check reproduces OutOfBoundsDatetime for the s and ms units without importing xarray; the same check passes with pandas 3.0.6. The follow-up commit only moves this PR's release note into the unreleased section. This is a known documentation-check failure, not a claim that the complete documentation build passes.

Checklist

• Closes Remaining cupy failures: cupy indexers on numpy objects, basic assignment, bool() #11739
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst

AI Disclosure

• This PR contains AI-generated content.
  • The AI-generated changes were tested by Codex on the remote workstation.
  • The contributor has reviewed the changes and reported no issues.

Tool: Codex. Task scope: address the CuPy indexing, assignment and scalar-bool failures reported in #11739, with regression tests and validation on the remote workstation. Codex prepared the six-file patch, ran the reported checks, prepared this description and submitted the PR at the contributor's explicit request. The contributor's own review confirmation is quoted above. No personal execution of the agent-run tests is claimed.

[headtr1ck]

Thanks! I ran this on a GPU (cupy 14.2.0): all tests in test_cupy.py and the cupy cases in test_duck_array_wrapping.py pass, while 10 of the new tests and test_idxmax/test_idxmin[True-cupy] fail on main, so this fixes everything from #11739.

The main thing I'd like to change is that the fix is quite specific to cupy, while the underlying problem is general: pandas can only be indexed with numpy indexers, so any non-numpy duck array indexer has to be converted at the pandas boundary.

• PandasIndex.isel already does this generically, but PandasIndexingAdapter._index_get only converts cupy arrays, and does it twice (key, then indexer.tuple for the multidimensional case). Converting the indexer once at the start of _index_get with the same rule as in PandasIndex.isel (is_duck_array(k) and not isinstance(k, np.ndarray) → to_numpy(k)) removes both the special case and the duplication. I tried that locally and the cupy tests still pass.
• Could you add a short comment to __bool__ on why extension arrays are special-cased? They have no __bool__, so bool() falls back to len(), and bool(pd.array([False])) would be True.
• Optional: NumpyIndexingAdapter._safe_setitem still reads array.flags.writeable, which cupy's flags don't have, so other ValueErrors from cupy assignments still turn into an AttributeError. getattr(array.flags, "writeable", True) would fix that.
• The whats-new entry can drop "while preserving pandas extension-array behavior", that's not a user-visible change.

The docs failure is the pandas 3.1 issue fixed in #11796, merging main should fix it.

[headtr1ck]

Thanks, all points are addressed! I reran the cupy tests on a GPU (cupy 14.2.0) after merging main: test_cupy.py and the cupy cases in test_duck_array_wrapping.py pass, as do the indexing, variable, dataarray, dataset and indexes tests.