[ty] fix panic with direct recursive type alias

[mtshiba]

Summary

I found that trying to use a directly recursive type alias like this would result in a stack overflow:

type Itself = Itself

def foo(Itself: Itself):
    x: Itself

This should be reported as an invalid type alias, but a stack overflow should be fixed immediately anyway. We now consider the value type to be Unknown in such cases.

Test Plan

New test case in mdtest\pep695_type_aliases.md

[astral-sh-bot]

Diagnostic diff on typing conformance tests

No changes detected when running ty on typing conformance tests ✅

[astral-sh-bot]

mypy_primer results

No ecosystem changes detected ✅

No memory usage changes detected ✅

[MichaReiser]

Does this fix also work for compound types like list[Itself]?

[carljm]

Those generally already work, without this PR: https://play.ty.dev/a65c2d75-e9c1-4a74-85e3-f96db2dd9c60

[carljm]

Another case that may be worth testing is multi-step direct recursion:

type A = B
type B = A

[mtshiba]

Hmm, unfortunately, it seems the panic still persists in the case of mutually recursive aliases:

type A = B
type B = A

def bar(a: B):
    x: a

Therefore, the measure is insufficient, and we should check the consequences of eagerly expanding a type alias value type before we first use of it.
This may be an issue that needs to be resolved in #20566.

[carljm]

Thank you!

[carljm]

IMO it's a bit odd to test that you can annotate a parameter as Itself, and then use that parameter again as an annotation. It may work due to how we treat Unknown in a type annotation, but I don't think this is something we want to commit to in a test.

I would probably just have done this instead:

Suggested change

	def foo(Itself: Itself):
	x: Itself
	def foo(x: Itself):
	reveal_type(x) # Unknown

Is there a particular reason why you think it's important to test this "annotated value as annotation" behavior?

[mtshiba]

In fact, in that example, no panic occurs, and the revealed type is Never.

https://play.ty.dev/4732e30a-301c-4753-b95f-aed34cf08585

Should the original code be placed in the corpus as a regression test?

[carljm]

Ah, I wondered if this part was necessary to reproduce the panic. That explains why I didn't catch this case in my previous work on recursive type aliases. Makes it much less likely people will run into this, but of course it should still be fixed.

I think it's OK for a test like this to be in mdtests, I would just expect some commentary (specifically about annotating with a value type) that "this is a very strange thing to do, but this is a regression test to ensure it doesn't panic"

[mtshiba]

Detecting mutually recursive invalid type aliases was harder than expected, and I decided to address it in #20566.

Related commits:
ac44afa
2a9b9a2

I will close this PR.