Generic type inferred wrong; curious difference betweer `aiter(f(x))` and `c = f(x); aiter(c)`

Bug Report

Consider this code (playground):

from typing import TypeVar, Generic, Callable, Coroutine, Any, AsyncIterator

_T = TypeVar("_T")

class StartedToken:
    pass

class ProducerContext(Generic[_T]):
    pass

async def coro_to_aiter(
    coroutine: Callable[[ProducerContext[_T]], Coroutine[Any, Any, None]]
) -> AsyncIterator[_T | StartedToken]:
    yield StartedToken()

async def coro(ctx: ProducerContext[int]) -> None:
    pass

# Split into two lines, no error:
c = coro_to_aiter(coro)  # no error
it = aiter(c)  # no error

# The same on one line -> mypy error:
it2 = aiter(coro_to_aiter(coro))  # error: Argument 1 to "coro_to_aiter" has incompatible type ...

Expected Behavior

I would expect mypy to not error on the it2 definition line. Moreover, regardless of whether mypy can handle these definitions, I would expect to see the same behavior (error or absence of error) in the "Split into two lines" case and the "on one line" case.

Actual Behavior

I get this error on the it2 line, and no errors on other lines:

main.py:24: error: Argument 1 to "coro_to_aiter" has incompatible type "Callable[[ProducerContext[int]], Coroutine[Any, Any, None]]"; expected "Callable[[ProducerContext[int | StartedToken]], Coroutine[Any, Any, None]]" [arg-type]

Your Environment

• Mypy version used: 1.6.1 (playground), current master
• Python version used: 3.11, 3.12

Start with the main.py reproducer from the issue, using the listed Python and mypy versions, and compare the split and nested aiter(coro_to_aiter(coro)) cases. Trace how generic type inference handles each call and add or update a focused regression test for the discrepancy. Done means both forms receive consistent, expected diagnostics.