Problems with parametric types and promotion

[nstarman]

Consider

from dataclasses import dataclass
from plum import promote, parametric, add_promotion_rule
from typing import Any

@parametric
@dataclass
class Class:
    a: int | float
    b: int | float

    @classmethod
    def __init_type_parameter__(self, a: type, b: type) -> tuple[type, type]:
        """Check whether the type parameters are valid."""
        return a, b

    @classmethod
    def __infer_type_parameter__(self, a: Any, b: Any) -> tuple[type, type]:
        """Inter the type parameter from the arguments."""
        print(a, b)
        return type(a), type(b)

x = Class(1, 2)  # Class[int, int]
y = Class(1.0, 2.0)  # Class[float, float]

plum appears to auto-register a self-promotion pathway, but parametric types produce subclasses, so the promotion doesn't cover the types.

promote(x, y)
# TypeError: No promotion rule for `__main__.Class[int, int]` and `__main__.Class[float, float]`.

If I register a generic promotion for Class to cover its subclasses then I hit the the MethodRedefinitionWarning

add_promotion_rule(Class, Class, Class)
promote(x, y)
# MethodRedefinitionWarning: `Method(function_name='_promotion_rule', ...
# (Class[int, int](a=1, b=2), Class[float, float](a=1.0, b=2.0))

There does not appear to be a way to register a generic dispatch for parametric types without incurring MethodRedefinitionWarning!

[wesselb]

@nstarman I believe this is due to this bit of the promotion mechanism, which, in this case, indeed defines a method that overwrites another.

I'm strongly considering whether we should make the redefinition warnings opt in, so e.g. one would write dispatch = Dispatcher(warn_redefinition=True).

[nstarman]

I think dispatch = Dispatcher(warn_redefinition=True) is a good idea!