Here's a reduction against github/master 63856d4:

protocol P {
    associatedtype A
    var a: A { get }
}
extension P where A == Int {
    var a: Int { return 0 }
}
struct X : P {  } // Compiles?

If this is working as intended, I think we made a mistake in the semantics of the generics system. It certainly doesn't match my mental model for how things are supposed to work. I would expect it if I had written associatedtype A = Int or added

extension P { typealias A = Int }

(which incidentally doesn't compile). To put it in English, I expect extension P where A == Int to be a conditional extension that only applies when the associated type is deduced to be `Int` by other means, not to influence deduction of the associated type on which it is itself conditional.

cc @huonw, @rjmccall

I agree with Dave that we shouldn't default anything in order to match a conditional conformance.

@swift-ci create

This is related to SR-7610

I linked as related to SR-5440, because it seems so similar to eg the following (even though it's not involving protocols, only constrained extensions):

This (invalid) program compiles successfully, because the compiler misinterprets the where clause:

struct X<T> {
    var a : A
}
extension X where T == Never {
    typealias A = Int
}
let x = X<Int>(a: 123)

Note that this isn't a conditional conformance. cc @DougGregor since it's an associated type inference thing.