dataclasses.replace: fall through to typeshed sig

[ikonst]

If the dataclasses plugin cannot determine a signature for dataclasses.replace, it should not report an error. The underlying typeshed signature will get a shot at verifying the type and reporting an error, and it would enable the following pattern (without typing replace's kwargs, though):

from dataclasses import is_dataclass, replace

def foo(x: object):
  if is_dataclass(x) and not isinstance(x, type):
#    ^^^^^^^^^^^^
#    this is a type-guard in the typeshed
#
    y = replace(x)

[sobolevn]

Generally looks good, just one question :)

[sobolevn]

I think that this can potentially introduce some new weak spots, but on other hand it can eventually work :)

[ikonst]

This is a false positive but it's due to #15974.

[sobolevn]

Do we need to test just if is_dataclass(x):?

[ikonst]

It's because type is an object, and is_dataclass on an object either means it's a dataclass instance or type.

I'll improve the test to clarify that.

[ikonst]

1a538d4

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

vision (https://github.com/pytorch/vision): typechecking got 1.37x slower (36.8s -> 50.5s)
(Performance measurements are based on a single noisy sample)