Unconstrained type parameter has stricter comparability rules than `extends unknown`

[DanielRosenwasser]

Fine in 4.7 and 4.6

function foo<T extends unknown>() {
 let x = {};
 x as T;
}

Fine in 4.6, fails in 4.7

function foo<T>() {
 let x = {};
 x as T;
}

[weswigham]

Weird. I wonder what comparability is doing to make this happen.

[DanielRosenwasser]

Probably this #32814

[weswigham]

So comparability, as a relation, is in a very bad way from a consistency perspective right now.

Take this, for example:

enum E { a, b, c }
 
declare var a: boolean;
declare var b: number;
declare var c: string;
declare var d: void;
declare var e: E;
declare var f: { a: string };
declare var g: any[];

function fn<T>(t: T) {
 var r8a1 = t < a;
 var r8a2 = t < b;
 var r8a3 = t < c;
 var r8a4 = t < d;
 var r8a5 = t < e;
 var r8a6 = t < f;
 var r8a7 = t < g;
 var r8b1 = a < t;
 var r8b2 = b < t;
 var r8b3 = c < t;
 var r8b4 = d < t;
 var r8b5 = e < t;
 var r8b6 = f < t;
 var r8b7 = g < t;
}

this is an except from our tests. All of these comparisons are forbidden (because instantiation may make the comparison forbidden, was our stated logic in the relevant issue). However, this

enum E { a, b, c }
 
declare var a: boolean;
declare var b: number;
declare var c: string;
declare var d: void;
declare var e: E;
declare var f: { a: string };
declare var g: any[];

function fn<T extends unknown>(t: T) {
 var r8a1 = t < a;
 var r8a2 = t < b;
 var r8a3 = t < c;
 var r8a4 = t < d;
 var r8a5 = t < e;
 var r8a6 = t < f;
 var r8a7 = t < g;
 var r8b1 = a < t;
 var r8b2 = b < t;
 var r8b3 = c < t;
 var r8b4 = d < t;
 var r8b5 = e < t;
 var r8b6 = f < t;
 var r8b7 = g < t;
}

is the same but with one small difference - an explicit constraint. This currently issues no errors on any comparison, despite ostensibly being identical. This is, perhaps obviously, terribly inconsistent.

The request in the OP of this thread, that

function foo<T>() {
 let x = {};
 x as T;
}

be allowed, implies that all of the comparisons I listed above should be allowed. It's certainly doable (and in fact done, implementation on-hand over in #48861). Conversely, if in presenting this inconsistency I've changed your mind and now you think all these comparisons should be forbidden, we need to add a lot of errors to constrained type parameters in comparability (which is going to be much more breaking)!