[ilevkivskyi]

Fixes #17326
Fixes #21739
Fixes #21985

There are scenarios where recursive protocols are genuinely undecidable (a.k.a. diverging). Consider the example in testDivergingProtocol: to check if C <: P[int] we need to first check if C <: P[list[int]] assuming C <: P[int], but to check that we need to first check if C <: P[list[list[int]]] etc. This (classic) algorithm never converges.

Note I set a relatively large limit of 40. After playing manually with this I found that some libraries can use up to ~25 assumptions in the stack, so I want to be on the safe side initially, we can optimize this later. To be clear, this should not cause any performance issues for libraries that don't crash currently.

Now on why do we do this only for recursive protocols and not for recursive type aliases:

• This comes up more often for protocols (FWIW I don't think there is a single issue about the same problem with recursive aliases)
• There is already a way to detect such problematic aliases early during semantic analysis, so we give a normal error, not a crash.
• Fixing this for recursive aliases is more tricky, since it will require to patch a lot of visitors involving the seen_aliases pattern.

cc @JukkaL

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[JukkaL]

Makes sense -- potential for very rare false negatives is better than crashing/hanging.

[quentin]

Many thanks!