type inference is fragile for protocols like SupportsKeysAndGetItem

Repro: https://mypy-play.net/?mypy=latest&python=3.11&gist=d2f8dfa1fdfda5c0c286f1a73f5e39eb

Pasted below for convenience:

import typing as t

from _typeshed import SupportsKeysAndGetItem


KT = t.TypeVar("KT")
VT = t.TypeVar("VT")


def version1(obj: SupportsKeysAndGetItem[KT, VT] | t.Iterable[tuple[KT, VT]]):
    if hasattr(obj, "keys") and hasattr(obj, "__getitem__"):
        reveal_type(obj)  # SupportsKeysAndGetItem[KT, VT] | Iterable[tuple[KT, VT]] (as expected)
    else:
        reveal_type(obj)  # narrowed to Iterable[tuple[KT, VT]] as expected


def version2(obj: SupportsKeysAndGetItem[KT, VT] | t.Iterable[tuple[KT, VT]]):
    if hasattr(obj, "keys") and hasattr(obj, "__getitem__") and (obj_getitem := obj.__getitem__):
        reveal_type(obj)  # SupportsKeysAndGetItem[KT, VT] | Iterable[tuple[KT, VT]] (as expected)
    else:
        reveal_type(obj)  # NOT narrowed

I would expect version 2 to behave the same as version 1. Is that an improvement mypy could reasonably make?

I searched for relevant docs and existing issues, but didn't find anything specifically about this. Apologies if I missed something.

Run the linked mypy-play repro and compare the reveal_type results in version1 and version2. Trace the type-narrowing logic for the combined hasattr checks and walrus assignment; done means version2 narrows the else branch consistently with version1, with regression coverage for the example.