Constrained TypeVars with multiple valid solutions

[JelleZijlstra]

Given this code:

from typing import TypeVar, SupportsInt, SupportsFloat

T = TypeVar("T", SupportsInt, SupportsFloat)
U = TypeVar("U", SupportsFloat, SupportsInt)


def f(x: T) -> T:
    raise NotImplementedError


def g(x: U) -> U:
    raise NotImplementedError


reveal_type(f(1.0))  # mypy, pyre, pyright: SupportsInt; pytype: float
reveal_type(g(1.0))  # mypy, pyre, pyright: SupportsFloat; pytype: float

A float is both a SupportsInt and a SupportsFloat, so both constraints on the TypeVar are valid solutions. But all of mypy, pyre, and pyright simply pick the first solution. (Pytype picks "float", which I don't think should be allowed.) I feel like inference should not depend on the order in which the constraints are defined, and in pyanalyze I fall back to Any if there are multiple valid solutions.

But mypy, pyre, and pyright are all happy to rely on the order of the constraints, so is it OK to do so?