Fix noUncheckedIndexedAccess with tuple rest types and generic index types

[andrewbranch]

(each commit fixes a separate bug)

Fixes #40657
Fixes #40666

[andrewbranch]

At this point, we don’t yet know whether the index type will be constrained to something that resolves to property signatures or index signatures, so we can’t return a union with undefined. But when we come back to do the deferred checking of the indexed access type we saved, we’ve lost the accessFlags that indicate that this is a position where undefined should be included in the type.

As I’m writing this, I’m realizing that this is probably broken for generic indexed access types that get used both for reading and for writing, so converting back to draft.

[andrewbranch]

The only easy way I could find to fix this was creating separate IndexedAccessTypes (and keying the cache) depending on whether it was an expression where --noUncheckedIndexedAccess could have an effect or not. So in effect, in

function<T extends string>(key: T) {
  obj[key] = undefined;
  return obj[key];
}

the two usages of obj[key] have two separate types. (This only happens when the object type and/or the index type is generic.) I think this is technically possible to avoid, but would involve a lot of inexpedient threading of accessFlags (or some parameter based on it) through a bunch of functions.

[DanielRosenwasser]

Maybe this should just be named shouldIncludeUndefined?

[andrewbranch]

I did that at first, but I felt like it was confusing because that’s only true if the type resolves to an index signature. Indexed access types often resolve to an actual property, where they should not include undefined. Happy to hear pithier suggestions that don’t imply the unconditional presence of undefined though.

[sandersn]

@andrewbranch is this ready to merge, or is it waiting on a review from @weswigham? I guess it's the latter but I want to make sure.

[andrewbranch]

Ryan suggested waiting on a review from @weswigham if possible.

[weswigham]

My comment is mostly stylistic - this looks OK implementation-wise (noUncheckedIndexedAccessCandidate is super wordy, though 😛 )

[andrewbranch]

@weswigham and @DanielRosenwasser both expressed distaste toward the name; I don't like it either, but I like it better than shouldIncludeUndefined for the reason I outlined here. Open to suggestions before merging.