Type hinting a union of callables does not work as hoped

[Hugovdberg]

I'm trying to implement a robust version of the builtin max function by performing a reduction on an iterable with a given initial value (which eventually will be curried using the expression package):

def max(
        it: Iterable[_a], 
        default: _a, 
        keyfunc: Callable[[_a], _SupportsLT] | Callable[[_a], _SupportsGT],
) -> _a:
    return functools.reduce(
        lambda acc, value: acc if keyfunc(acc) > keyfunc(value) else value, 
        it, 
        default,
    )

where _SupportsLT and _SupportsGT are identical to the protocols defined in the stdlib typeshed.

Now the problem is that Pylance raises a reportOperatorIssue on the keyfunc(acc) > keyfunc(value) part. This is because keyfunc(acc) has type SupportsLT | _SupportsGT, and the same for keyfunc(value), and _SupportsLT and _SupportsGT cannot be compared. However, given the fact that the keyfunc either always returns a _SupportsLT or always a _SupportsGT the reported error condition would never occur.

Is there a way to modify the type hints such that Pylance can infer that the comparison will always succeed?

[erictraut]

If keyfunc(acc) returns a value that supports __lt__ and keyfunc(value) returns a value that supports __gt__, the function would fail. The expression keyfunc(acc) > keyfunc(value) would generate an exception at runtime. So I think pyright is correct in telling you that there's an error here.

[Hugovdberg]

But in my mind the type hint states that keyfunc should either be a function that turns all a's in to something that supports __lt__ or turns all a's into something that supports __gt__ (or both __lt__ and __gt__), but not some a's into __lt__ and other a's in __gt__ supporting types. Since acc and value always are of the same type, the result of keyfunc should also always be of the same type, right?

Is it correct that the current implementation interprets Callable[[_a], _SupportsLT] | Callable[[_a], _SupportsGT] as equal to Callable[[_a], _SupportsLT | _SupportsGT]? Because that seems to me what's happening here.