This patches a surprising difference in the inference behavior in respect to return types:

declare function test1<T>(value: T): T;

const result1 = test1("foo");
//    ^? const result1: "foo"

declare function test2<T>(value: T): T extends string ? T : never;

const result2 = test2("foo");
//    ^? const result2: string

The rule is that T has to appear in a top-level position in the return type for its literalness to be preserved. From the user's point of view ,it is in test2's top-level position in the return type. But since T in that truthy branch is seen as a substitution type for the compiler it fails to recognize it.