Laziness and TMC are fundamentally different from partial instantiation in that they are implementation details, mostly/wholly invisible to the language semantics themselves.

A key aspect of partial instantiation is that the "holes" may be filled in by a semantically unrelated piece of code, which is not the case for either laziness or TMC (wherein the contents data structure must be defined in one place, even if the implementation does not evaluate it immediately).

(I don't know Koka so I can't speak to that.)

Also, you can do lazy initialization in any language that has functions by passing a getter function around instead of a value.

I recently had need for this to implement validation for recursive data structures in TypeScript. It depends on support for forward references in the body of a function. The tricky bit is ensuring that the getter isn’t called until the cycle is completed by defining the forward reference’s target. The type system doesn’t help; you just have to write the implementation so it doesn’t call the getter.

I don't think Haskell can do this, can have a growable linked list for example. 'last a' is 'last a', regardless what is between them (modulo shadowing and such).

And I suspect that Prolog's Partial Instantiation is, while not mutating data, but it is mutating references somewhere