`xr.decode_cf()` fails to replace `FillValue` with `np.nan` with `numpy >= 2.0` when data array dtype is `np.float32`

What happened?

Since numpy >= 2.0, xr.decode_cf() fails to replace FillValue with np.nan when:

• the data array dtype is np.float32
• the FillValue attribute is of type float.

The issue does not occur for data arrays with dtype np.float64.

This bug should not affect decoding of dataset opened with open_dataset because since numpy >= 2.0, numeric values in the attributes are now directly stored as np.int<...> or np.float32 types instead of the previously used int or float types.

To keep my code compatible with all numpy versions I currently use:

if version.parse(np.__version__) >= version.parse("2.0.0"):
    vars_and_coords = list(ds.data_vars) + list(ds.coords)
    for var in vars_and_coords:
        if "_FillValue" in ds[var].attrs:
            ds[var].attrs["_FillValue"] = ds[var].data.dtype.type(ds[var].attrs["_FillValue"])

What did you expect to happen?

xr.decode_cf() to replace fillvalues with np.nan also when the FillValue attribute is of type float.

Minimal Complete Verifiable Example

import numpy as np
import xarray as xr
 
fill_value = -99.9 
data = np.ones((2,3)) * fill_value
data = data.astype(np.float32) # The error does not occur with float64 !

ds = xr.DataArray(data, dims=["x", "y"]).to_dataset(name="var")
ds["var"].attrs["_FillValue"] = fill_value # float 
xr.decode_cf(ds)["var"].data # do not replace -99.9 values with np.nan
ds["var"].attrs["_FillValue"] = np.float32(fill_value) # np.float 
xr.decode_cf(ds)["var"].data # do replace -99.9 values with np.nan

MVCE confirmation

• Minimal example — the example is as focused as reasonably possible to demonstrate the underlying issue in xarray.
• Complete example — the example is self-contained, including all data and the text of any traceback.
• Verifiable example — the example copy & pastes into an IPython prompt or Binder notebook, returning the result.
• New issue — a search of GitHub Issues suggests this is not a duplicate.
• Recent environment — the issue occurs with the latest version of xarray and its dependencies.

Environment

Start by running the provided minimal example against xr.decode_cf() with NumPy 2.0 or newer, comparing float32 and float64 data and Python float versus np.float32 fill values. Trace the decode_cf entry point and its fill-value handling; done means the float32 array replaces matching fill values with np.nan while existing float64 behavior remains unchanged.