Inferred attribute type feeds back as call context, rejecting its own defining assignment

Bug Report

Assigning a generic call returning K | None to a previously undeclared
self. attribute spuriously fails [arg-type] on that same assignment. The
attribute's inferred type (str | None) apparently comes back as call context
on a later pass, solving K = str | None; the invariant Mapping key then
rejects dict[str, str].

The same expression passes with no context, an annotated local, an annotated
attribute, or a reassigned local (whose second assignment is likewise checked
against its own inferred str | None) — the optional-stripping special case
seems skipped on the inferred-attribute path alone. A sync version (no
async/await) also passes, pointing at the deferred re-check await
triggers.

To Reproduce

from collections.abc import Mapping


async def select_one[K](options: Mapping[K, str]) -> K | None: ...


class ViaInferredAttr:
    async def m(self) -> None:
        cases = {"a": "x"}
        self._done = await select_one(cases)  # error (spurious)


async def via_local() -> None:
    cases = {"a": "x"}
    done = await select_one(cases)  # ok
    done = await select_one(cases)  # ok (context = inferred str | None)
    done2: str | None = await select_one(cases)  # ok


class ViaAnnotatedAttr:
    _done: str | None

    async def m(self) -> None:
        cases = {"a": "x"}
        self._done = await select_one(cases)  # ok

Expected Behavior

All call sites typecheck, with K = str.

Actual Behavior

repro.py:10: error: Argument 1 to "select_one" has incompatible type "dict[str, str]"; expected "Mapping[str | None, str]"  [arg-type]

The str | None key shows the return's | None leaked into K.

Your Environment

• mypy 1.20.2 (compiled), no flags
• Python 3.13

Related: #5874 (over-eager outer context with unions), #19304 (assignment target
changes results). New here: the context is self-inferred — no annotation exists
anywhere — and explicit spellings of the same type behave differently.

Start with the reproducer in the issue and trace the deferred async re-check for an assignment to an undeclared self attribute. Compare that path with the passing local and annotated-attribute cases, focusing on how inferred return context reaches generic argument inference. Done means the reproducer passes with K inferred as str and a regression test covers the behavior.