When calculating inferred infer type constraints, replace conditionals…

[weswigham]

…within mapped type templates with their constraints.

Fixes #44143

[weswigham]

This comment block is informative for whomever comes back to this in the future - I looked at two other ways to solve the problem (the conditional in the template type executing differently than the mapped type around it would imply when instantiated in the above fashion), but both ran against other issues in the checker. If one of the other's root issues get fixed, I think they'd be much more elegant constructions to return here than this conditional-constraint-walking logic (ideally the first of the two, since it much more closely matches the structure traversed).

[weswigham]

Fun fact - this test already showed the issue, we just weren't looking for it in this test. (Instead we were looking for a constraint error on a different line.)

[RyanCavanaugh]

@ahejlsberg 🔔

[ahejlsberg]

I'm somewhat puzzled by the logic that attempts to add constraints to infer type parameters in a very narrow scenario (inferring from a mapped type to another mapped type where the infer type parameter occurs in the template type). It was added in #43649 and this PR now makes further modifications to it, but I don't see why any of it is needed. Indeed, if I simply delete the code from #43649, everything works as expected, including the broken test in inferConditionalConstraintMappedMember.ts.

[Andarist]

@ahejlsberg i stumbled upon this one while looking into something else - removing the code from #43649 indeed fixes the expected type result (as this PR does, from never to 'foo' | 'bar') but removing it brings back the assignability error that was fixed by #43649

    export type RemoveIdxSgn<T> = Pick<T, KeysWithoutStringIndex<T>>
                                          ~~~~~~~~~~~~~~~~~~~~~~~~~
!!! error TS2344: Type 'KeysWithoutStringIndex<T>' does not satisfy the constraint 'keyof T'.
!!! error TS2344:   Type 'unknown' is not assignable to type 'keyof T'.

[Andarist]

There are currently (at least) 3 issues open related to this: #46020 #44143 #44351

All of them would get fixed if #43649 would get reverted (as in, that PR broke them). "Known keys" would get fixed in all of them if this PR here would get merged but the third issue also suffers from index access error and that is not fixed by this PR here.

Note for other reviewers: the main problem here is that keyof doesn't exactly get the same keys as K within mapped types (K within mapped types can be used to iterate over each key, including index signatures whereas keyof "reduces" the union of possible keys). So when the mapped type from the checkType gets unwrapped away and substituted with the constraint of the template (which is the conditional type referring to K) then we essentially "lose" the information for this case because K gets replaced with keyof T (and that gets reduced from 'knownKey1' | 'knownKey2' | string | number to string | number).

[sandersn]

@weswigham This has been open for a couple of years. If this isn't the right solution, should we close it?
Is it possible to fix the error that @Andarist mentions some other way?

Either way, this fix is probably minor enough to wait until Corsa has overtaken Strada.