This error is caused by a limitation in mypy's type inference system, and won't occur with other typecheckers like pyright or basedpyright. See #27957 for a (rather lengthy) discussion on this in detail.

Luckily, there's an easy workaround:

import numpy as np
import numpy.typing as npt

components_proper = np.random.random([3, 3]) - 0.5
components_improper = np.random.random([3, 3]) - 0.5

in_SST_: npt.NDArray[np.bool] | np.bool
if np.random.randint(1, 3) > 1:
    in_SST_ = np.logical_or(
        np.all(components_proper >= 0.0, axis=-1),
        np.all(components_improper >= 0.0, axis=-1),
    )
else:
    in_SST_ = np.all(components_proper >= 0.0, axis=-1)

This helps mypy to understand what the common denominator of in_SST_ is, because it isn't able to figure that it by itself.
If possible for you, you might also want to consider switching to pyright or basedpyright, which don't have this issue.

But even so, you're right in saying that the return type of logical_or is too narrow. Because, like np.all, its return type should also include to possibility for scalars, not only arrays.
But as you can see in

Somehow this has gone by unnoticed for 4 years, so thanks for reporting this!

Some related issues:

• Typing support for shapes #16544
• TYP: Overlapping overloads with incompatible return types #27032
• TYP: regression between 2.1.3 and 2.2.0 (mypy only) #27957

FYI, there's a newer v1.14.1 mypy release available

@jorenham
thanks for the quick reply and the detailed explanation.

Actually, I think everything is correct for the typehint in line 190, at least for the case I had in mind. If only one of the arguments of np.logical_and is a scalar, the return value is an array. Only two scalars (the other @overload) will give a scalar. The problem is just that mypy apparently infers the type from the first assignment instead of checking all assignments for a common denominator.

The bug is rather subtle:

import numpy as np
import numpy.typing as npt

a: npt.NDArray[np.float64] | np.float64
b: npt.NDArray[np.float64] | np.float64
out = np.logical_or(a, b)
reveal_type(out)  # npt.NDArray[Any]

So out is inferred as array, even though both inputs could also be a scalar.