Fix type inference in pattern matching by positional argument

[sobolevn]

Oh, this was very interesting.
During the debug sessions I learned a lot about how pattern matching and its analysis do work.

But, the problem was that expand_type did not preserve .last_known_value for some reason.
I used .copy_modified to preserve everything we know about Instance.

However, I expect that some tests might fail now. This code even has something similar in TODO some lines above:

mypy/mypy/expandtype.py

Lines 144 to 148 in 88aed94

	repl = self.variables.get(t.id, t)
	if isinstance(repl, ProperType) and isinstance(repl, Instance):
	# TODO: do we really need to do this?
	# If I try to remove this special-casing ~40 tests fail on reveal_type().
	return repl.copy_modified(last_known_value=None)

Let's see what will happen.

Closes #13612

[sobolevn]

I am quite surprised!

[ariebovenberg]

@sobolevn a good surprise indeed! Tests look comprehensive enough.

I wonder, does it also run correctly with inheritance? (Note the extra a: Base[str] = ... alias)

class Base(Generic[T]):
    ...

@dataclass  # problem also occurs with non-dataclass equivalent
class A(Base[T]):
    x: T


a: Base[str] = A("foo")
reveal_type(a)  # Base[str] (correct)

match a:
    case A(b):
        reveal_type(b)  # Any (incorrect! Should be builtins.str)

It might be a good addition to the tests, because inheritance here breaks dataclass which did work without it.

[sobolevn]

I will add a test case for subclasses and dataclasses. Thanks!

[sobolevn]

Ok, this indeed does not work for Generic subclasses. But, I belive that it is something different.

Let's open a new issue for it. Can you, @ariebovenberg? 🙂

[github-actions]

According to mypy_primer, this change has no effect on the checked open source code. 🤖🎉

[sobolevn]

@thomkeh hi, can you please share any context on your commit? Is there something wrong?

[tmke8]

Oh sorry, I clicked the "revert" button accidentally! I didn't know this was possible.

Everything is fine as far as I know.