false negative: `int` and `bool` don't implement `.hex()`

[jorenham]

Bug Report

builtins.float and builtins.int shouldn't be assignable to a Protocol implementing hex: () -> str, because that would be type-unsafe:

>>> int().hex()
Traceback (most recent call last):
  File "<python-input-0>", line 1, in <module>
    int().hex()
    ^^^^^^^^^
AttributeError: 'int' object has no attribute 'hex'
>>> bool().hex()
Traceback (most recent call last):
  File "<python-input-1>", line 1, in <module>
    bool().hex()
    ^^^^^^^^^^
AttributeError: 'bool' object has no attribute 'hex'

Mypy accepts this regardless.

To Reproduce

from typing import Protocol

class CanHex(Protocol):
    def hex(self, /) -> str: ...

True negative:

float_has_hex: CanHex = 3.14   # accepted

True positives:

object_doesnt_have_hex: CanHex = object()  # rejected: [assignment]
complex_doesnt_have_hex: CanHex = 1j  # rejected: [assignment]

False negatives:

int_doesnt_have_hex: CanHex = 666  # accepted?
bool_clearly_also_doesnt_have_hex: CanHex = False  # accepted??

https://mypy-play.net/?mypy=latest&python=3.12&gist=8ac210b6e647983b5c165af053ccd7d1

This isn't limited to .hex(), but applies to all methods that are implemented by builtins.float but not by builtins.int.

---

I consider this a high-priority issue, because of the following reasons:

1. It's type-unsafe, even in the most explicit of situations: def tohex(x: CanHex): return x.hex()
2. This issue, in combination with float and complex forcibly change invariant types to covariant #18257, makes it impossible to distinguish between int and float input using static typing.

The second reason in particular is a big problem for NumPy, because it forces us to annotate e.g. np.dtype(complex) as np.dtype[np.complex128 | np.float64 | np.int_ | np.bool], and np.array(3.14) as np.ndarray[tuple[()], np.dtype[np.float64 | np.int_ | np.bool]].
So I'm sure you can see how this is confusing for users, especially if you consider that usually tends to propagate to other dtype- and array-types.

We can't use the "Any trick" either, because that'd be problematic for functions that overload on dtype, of which there are many.

[tyralla]

Thank you for the report. I am not at all familiar with the promotion topic but dared to fix it. Let's see if it works...