Motivation

• Prevent misclassification of dtype-specifier strings/types as pandas extension dtypes and preserve correct NumPy-style promotion for explicit dtype inputs.
• Preserve and return pandas ExtensionDtype results (e.g., CategoricalDtype) where appropriate and avoid unwanted promotion to object.
• Fix downstream typing/usage sites that assume result_type returns np.dtype and ensure string/bytes promotion matches pandas semantics for indexing and alignment.

Description

• Refined xarray.core.dtypes.result_type to detect extension-dtype inputs only for actual extension types and not for dtype-specifier str/bytes/type objects, and to return ExtensionDtype when appropriate.
• Added a fast-path that collects explicit dtype-like inputs (np.dtype, type, str/bytes specifiers), runs maybe_promote_to_variable_width on them, and calls np.result_type to preserve fixed/variable-width string semantics.
• Updated logic for should_return_str_or_bytes so extension-array dtype detection excludes plain type objects.
• Improved union_unordered_categorical_and_scalar in xarray.core.extension_array to concatenate category Index objects and produce a unique pd.CategoricalDtype (preserving order and uniqueness).
• Guarded maybe_coerce_to_str in xarray/core/utils.py to only check .kind for actual np.dtype instances.
• Added a cast(np.dtype, ...) where result_type is used in a sparse conversion path (xarray/namedarray/core.py) to satisfy typing when a concrete np.dtype is required.
• Minor test updates: tighten typing in xarray/tests/test_dtypes.py with cast, and use pd.NA in categorical-null tests in xarray/tests/test_dataarray.py.

Testing

• Ran python -m pytest -c /dev/null xarray/tests/test_dtypes.py -k "result_type" and it passed (19 passed, 1 skipped).

Codex Task