Arg-type warning calling dataclasses.replace() with **dict for keyword arguments.

[flaviovs]

Bug Report

Calling dataclasses.replace() using a dict for dynamic arguments issues arg-type warnings.

To Reproduce

from dataclasses import dataclass, replace

@dataclass
class Foo:
    bar: str
    zee: int

obj1 = Foo(bar='abc', zee=1)

kwargs = {'bar': 'xyz', 'zee': 9}
obj2 = replace(obj1, **kwargs)

print(obj1)
print(obj2)

Program output:

$ python /tmp/test.py 
Foo(bar='abc', zee=1)
Foo(bar='xyz', zee=9)

Expected Behavior

No warnings from mypy.

Actual Behavior

$ mypy /tmp/test.py 
/tmp/test.py:13: error: Argument 2 to "replace" of "Foo" has incompatible type "**dict[str, object]"; expected "str"  [arg-type]
/tmp/test.py:13: error: Argument 2 to "replace" of "Foo" has incompatible type "**dict[str, object]"; expected "int"  [arg-type]

$ mypy --version
mypy 1.5.1 (compiled: yes)

Apparently Mypy is trying to match each dict element type to class attributes (which is good), but not considering the fact that the 2nd parameter of replace() is a var-keyword therefore **dict is perfectly fine.

Your Environment

$ python --version
Python 3.9.2

$ mypy --version
mypy 1.5.1 (compiled: yes)

[sobolevn]

cc @ikonst

[ikonst]

We don't do custom arg validation for replace, we return a signature and have mypy validate it as it does with any other. The same thing happens if you take dataclasses out of it:
https://mypy-play.net/?mypy=latest&python=3.11&gist=ec706ef461064cf9fb54d8ab8df8d5b5

We have plenty of dups: #5382, #15317 - more can be found.

(Curiously pyright seems to handle it correctly, not sure what it does differently.)

[ikonst]

OK, I was too quick to decide pyright handles it more correctly.

The difference is in Inference of List, Set, and Dict Expressions.

mypy infers {'bar': 'xyz', 'zee': 9} as dict[str, object], pyright infers as dict[str, Unknown], the latter is somewhat akin to Any. For example, this will be a false positive on pyright:

def f(*, a: int, b: str) -> None:
    return

f(**{'a': 42, 'b': True})

[AlexWaygood]

Yeah, I agree with @ikonst that this is really just a dupe of #5382