binary(): numpy/dask return input dtype while cupy returns float32

[brendancol]

binary() returns a different output dtype depending on the array backend. The numpy and dask+numpy paths return the input dtype; the cupy and dask+cupy paths always return float32. Switching backends in a pipeline silently changes the container dtype for the same values.

Reproduce

import numpy as np, xarray as xr, dask.array as da, cupy
from xrspatial.classify import binary

data = np.array([[np.nan, 1., 2.], [3., 4., np.inf]], dtype=np.float64)

print(binary(xr.DataArray(data), [1, 2]).dtype)                       # float64
print(binary(xr.DataArray(da.from_array(data)), [1, 2]).dtype)        # float64
print(binary(xr.DataArray(cupy.asarray(data)), [1, 2]).data.dtype)    # float32

Cause

_cpu_binary allocates its output with dtype=data.dtype, while _run_cupy_binary uses dtype='f4'.

Why it matters

The docstring example shows float32 output, and every other classifier in classify.py (reclassify, quantile, natural_breaks, and so on) emits float32 on all four backends. binary is the only one that doesn't. For integer input the numpy path returns an integer dtype, which can't hold the NaN sentinel the docstring promises for non-finite cells.

Fix

Allocate the numpy output as float32 so the four backends agree and the NaN sentinel is always representable.

Found by the metadata-propagation sweep on classify.

[brendancol]

Resolved on main. The float32 fix landed across two merged PRs:

• Make binary() output float32 on all backends #3514 — _cpu_binary now allocates dtype=np.float32, so the numpy and dask+numpy paths match the cupy backends and the NaN sentinel is always representable.
• binary(): type the dask+cupy meta so the lazy dtype matches float32 (#3508) #3513 — types the dask+cupy map_blocks meta to f4 so the lazy dtype matches the computed result.

All four backends now return float32. Covered by test_binary_output_dtype_float32 and test_binary_dask_output_dtype_float32 in xrspatial/tests/test_classify.py, both passing.

Closing as completed.