There's no hard guarantee that freeing an object runs in constant time. Reference counting tends to provide more predictable performance, but I don't think the phrase "hard guarantee" is appropriate.
You're conflating a bunch of things, but the biggest one here is that "hard performance guarantee" is not a synonym for "constant-time performance". Nevertheless, feel free to substitute your own preferred vocabulary. I'm just trying to get the underlying point across about how different GC and RC fundamentally are.
With optimized RC (which deffer deallocation) you have no "constant-time performance guarantee", because you don't know at which time the deffered deallocation will happen... And all modern RC implementations are heavily optimized - especially in Apple's ecosystem...
Yeah I think generally what people mean here is, "if I'm aware that freeing objects takes time, I expect to spend that time when I free the objects, not at some undetermined point in the future". That's not the case if you do any kind of deferring, at which point you're pretty close to a "classic" GC system.