No love for NotImplemented?

When overriding binary special methods (e.g., __eq__, etc.), returning NotImplemented is a common convention, and generally omitted from returned types. But omitting it doesn’t agree with poor old bear:

from beartype import beartype

class Sneetch:
    @beartype
    def __init__(self, star: bool):
        self._upon_thars = star

    @beartype
    def __eq__(self, other) -> bool:
        if isinstance(other, Sneetch):
            return True  # Sylvester McMonkey McBean can pound sand
        else:
            return super().__eq__(other)

assert Sneetch(True) == Sneetch(False)  # all good
assert (Sneetch(True) == 42) is False  # boom

This will result in the following:

beartype.roar.BeartypeCallHintPepReturnException: @beartyped Sneetch.__eq__() return "NotImplemented" violates type hint <class 'bool'>, as "NotImplemented" not bool.

Making NotImplemented an explicit part of the return type is apparently hard (see python/mypy#4791).

I’m not sure what the right play is here. One could probably make a pretty good case that one should omit the @beartype decorator and lean on the internals (in which case, this should be closed as not-a-bug). But I thought I’d raise it anyway to get another opinion.

[posita]

… You’re in luck… again.

For some definition of luck, perhaps. Or maybe the emphasis belongs on recognizing that it was exactly one of us who was lucky here.

Mypy does exactly what everyone wants.

It may be but a small consolation that this irony was not lost on me.

That means I now need to ~disembowel~ refactor beartype to do exactly the same thing. Global happiness ensues.

A part of me celebrates resolution. After reading 57e5e99, however, another part wonders with great trepidation, “What have I wrought?”

May we all draw the courage to persevere from some sliver of hope that somehow the cure turns out not to be worse than the disease.

[leycec]

Brilliant! I hadn’t realized Mypy also accepted non-boolean binary dunder methods, although that’s painfully obvious in hindsight.

I also realized there’s a related issue with our current approach: it doesn’t permit unannotated returns. Interestingly, mypy does permit that and is thus totally down with this wacky tabacky:

class CodeGoneWild(class):
    def __add__(self, other):   # <-- this is fine, bro
        return (
            42 if isinstance(other, CodeGoneWild) else
            NotImplemented
        )

So, we also need to explicitly handle the case of binary dunder methods lacking return annotations. The best place for that “fun” kludge is probably in the following if branch in the body of the private beartype._decor._code.codemain._code_check_return() function:

    # If this return is unannotated, generate code calling this callable
    # unchecked and returning this value from this wrapper.
    if hint is _RETURN_HINT_EMPTY:
        # >>>> DO SOMETHING MAGICAL AND RISKY HERE <<<<
        func_wrapper_code = CODE_RETURN_UNCHECKED

Both of us deserve a well-deserved break and it’s not even Monday. 😭