I agree with your statement that GUID scheme scalling universaly is a new kind of utopia.
My view point is not directly linked to application of unique ids in RFF though. I believe this is the case for a larger scope of which the semantic web is one part...
I believe you have layed out sensible arguments in your post. If we now take the bad scalability of GUID schemes as a real world's fact, I can conclude that the real world will ultimately rely on a complex mesh of separate GUID schemes (we already have the premices of it, right ;)
We would need some ways to reconcile these schemes where they intersect. What are your views on how this kind of reconcilation can be modeled with the current semantic tools?