Enable --allow-redefinition-new in self-check

[ilevkivskyi]

Now that the performance issues are fixed, I think this is ready. I also undo previously added workarounds.

cc @JukkaL @hauntsaninja

[ilevkivskyi]

Oh well, now there is a recursion error in mypyc, last time I tried this few days ago it worked, so must be something new.

[ilevkivskyi]

The recursion error happens here:

        elif isinstance(typ, PartialType):
            assert typ.var.type is not None
            return self.type_to_rtype(typ.var.type)

so it looks like we somehow got a PartialType that points to itself, LOL.

[ilevkivskyi]

Pushing a fix in a separate PR #20879

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[ilevkivskyi]

Minimal repro for the new issue:

[case testFooBar]
# flags: --allow-redefinition-new --local-partial-types
from typing import Optional

def foo() -> None:
    name: Optional[str] = None
    name = None
    for i in [1, 2, 3]:
        reveal_type(name)  # N: Revealed type is "None | builtins.str"
        if bool():
            name = "a"

This is caused by my binder PR #20862. I guess my understanding of the binder logic was wrong: current_value = None means no narrowing known in current frame, not in all enclosing frames. It is also surprising that this only breaks with --allow-redefinition-new while there is an explicit annotation. Probably bind_all also does something bad.

[ilevkivskyi]

No, I actually checked .get() looks in all frames, so far this is a mystery.

[ilevkivskyi]

Oh

                    changed = declaration_type is not None and not is_same_type(
                        type, declaration_type
                    )

should be

                   changed = changed or declaration_type is not None and not is_same_type(
                        type, declaration_type
                   )

pop didn't change because of i 🤦