Reconsider constraints involving parameter specifications

[A5rocks]

• Fixes Decorator factory function typing  #15037
• Fixes Regression in 1.2.0: callback protocol with Callable and Concatenate #15065
• Fixes (🐞) Silent crash regression in 1.2 with complex Concatenate return value #15073
• Fixes Maximum recursion depth when using ParamSpec/Concatenate #15388
• Fixes (🐞) 1.2 regression invalid error and bad message with Concatenate return value #15086

Yet another part of #14903 that's finally been extracted!

[A5rocks]

Mypy primer diff looks right!

[hauntsaninja]

one quick comment

[A5rocks]

Fake error

Decently close to coming up with a reliable way to trigger this logic, but funnily enough I found a paramspec bug 👻

Just marking it here before I forget:

from typing import Generic, Callable, Union
from typing_extensions import ParamSpec, Concatenate

P = ParamSpec("P")
P1 = ParamSpec("P1")

class Tester(Generic[P1]):
    @staticmethod
    def test_something(x: Callable[P, None]) -> Tester[P]: ...  # type: ignore[empty-body]

    @staticmethod
    def test_concatenate(x: Callable[P, None]) -> Tester[Concatenate[int, P]]: ... # type: ignore[empty-body]

def func(
    action: Callable[P, None],
) -> None:
    job = Tester.test_concatenate(action) if True else Tester.test_something(action)
    reveal_type(job)

err, actually i'm decently sure that's just inference being annoying :(

[A5rocks]

I'm not too sure why this was this way before and if this is entirely bad behavior. I added this (and updates to hashing and equality) because paramspec types have extra state that is important: their prefix (which Concatenate adds)

[github-actions]

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

discord.py (https://github.com/Rapptz/discord.py)
+ discord/ext/commands/core.py:1729: error: Overloaded function signatures 1 and 2 overlap with incompatible return types  [misc]
+ discord/ext/commands/core.py:1799: error: Overloaded function signatures 1 and 2 overlap with incompatible return types  [misc]

[A5rocks]

Hi, bumping for @hauntsaninja or maybe @JukkaL.

[hauntsaninja]

Thanks, looks great! Might be worth a follow up PR to see if we can deduplicate a little bit or factor some stuff out of visit_instance

[hauntsaninja]

nit: stray print

[A5rocks]

Oh oops, lol.

Err wait, IIRC that's in order to make any test failures easier to check (it'll only show stdout on test failure).

[ilevkivskyi]

I am going to revert this. IMO this is just a bunch of more hacks to compensate for older hacks. Current ParamSpec support is quite foundationally broken. That needs to be addressed first.

[A5rocks]

I do think this isn't necessarily hacky -- I basically updated the code directly based on how I expect things to work out.... but I understand that ParamSpec support as a whole is hacky so... Makes sense

(I'm still surprised that this solved any issues at all; I made this PR because things were being weird)

[ilevkivskyi]

Yes, right, this doesn't really adds new hacks. I think I have an idea how to make ParamSpec a bit more principled. I am going to include some part of cleanups in coming inference PR (just because I was touching related code). Then I will make a separate second PR. I see you are interested in this topic. I will keep you posted on Discord.