This looks pretty good to me, but would like @zygoloid to double check the scope rules here to make sure I'm not missing a problem.

Spot checking the changes to the tests, they all seem like pretty nice improvements to the SemIR output.

The .param change is a clear improvement.

I think the scope change is also an improvement -- it certainly makes the behavior more consistent between return objects and other names in the decl block. I'm slightly on the fence because this is going to lead to more name collisions in the case of redeclarations, but probably the biggest question is which scope is less surprising, and I think that using the entity scope rather than the declaration instruction's scope is better on that basis.

For what it's worth, my thinking had been that for a case like

fn Outer() {
  let template T:! type = i32;
  fn Inner() -> T { return 0; }
}

... it'd make more sense for the reference to T to be in the scope of Outer (so that we don't have references into a function body scope from outside, and because that's the scope in which the expression is notionally evaluated). But I think a reference here from the scope of Inner to the scope of Outer isn't too much of a surprise, and we'll see such references if the body of Inner names T anyway.