[ty] Fallback to Unknown if no type is stored for an expression

[carljm]

Summary

See discussion at https://github.com/astral-sh/ruff/pull/19478/files#r2223870292

Fixes astral-sh/ty#865

Test Plan

Added one mdtest for invalid Callable annotation; removed pull-types: skip from that test file.

Co-authored-by: lipefree [email protected]

[github-actions]

mypy_primer results

No ecosystem changes detected ✅
No memory usage changes detected ✅

[carljm]

I don't know what's up with "CodSpeed Macro Runners" but I don't think it's related to this PR.

[dcreager]

👍 I like this. There are definitely places where we're currently jumping through extra hoops to infer a type for every expression, but we can simplify those as needed moving forward.

This also opens the door for storing separate value form and type form types for each expression, I think? That would in turn let us simplify parts of the call binding machinery.

[dhruvmanila]

This also resolves astral-sh/ty#865 (I tested this PR locally on the examples mentioned in that issue), it might be useful to add a couple of test cases from that issue.

[MichaReiser]

I think it's worth a try and we can always come back if it makes tracking down bugs harder

What I would find useful is to add least log a tracing debug message if we take the fallback branch

[sharkdp]

What I would find useful is to add least log a tracing debug message if we take the fallback branch

I thought about that as well. In https://github.com/astral-sh/ruff/pull/19478/files#r2223870292, I suggested that we might only want to do that in valid code (and in this case, it could even be an error?). Because otherwise, if the Unknown fallback is now our preferred way of dealing with AST nodes inside invalid code, we would see a lot of these tracing messages (in the LSP), e.g. when the user is currently typing a annotation, but it's not yet valid?

[AlexWaygood]

I agree that some logging for the fallback case might be good, but also that it might be hard to stop that from being very noisy in an IDE context. Even if you only emit the logs when there's valid syntax, there are lots of situations where you might have valid syntax but an invalid type expression when you're halfway through typing an annotation

[MichaReiser]

If the Unknown fallback is now our preferred way of dealing with AST nodes inside invalid code, we would see a lot of these tracing messages (in the LSP), e.g. when the user is currently typing a annotation, but it's not yet valid?

That's a fair concern. I don't think it should impact most users because the LSP only shows info level logs by default.

[carljm]

The tracing feels likely to be quite noisy to me, I would rather try without it -- if anyone finds themselves debugging a case like this and the tracing would have helped, we can add it.