last word on RDF serialization and QNames?

Everyone,

What's the last word on RDF serialization and QNames---has this been 
settled, and if so, what's the authoritative document? I've found 
numerous discussions, but nothing definitive.

The problem, as everyone probably knows, is that once an RDF+XML 
document is processed, all qnames are collapsed into static URIs. That's 
fine for RDF interpretation, but rough for re-serialization.

I've read a document somewhere that describes how to guess at the 
original namespace+localname serialization given a URI (does anyone 
remember where to find this document?), but this doesn't work for some 
combinations such as http://www.w3.org/1999/xlinkhref (which would yield 
(http://www.w3.org/1999/, xlinkhref).

More worrisome is an RDF graph that contains the URI 
"example:mynamespace-myproperty". I can't think of a way to serialize 
this---if serialized with the null namespace and a local name of 
example:mynamespace-myproperty, a namespace-aware XML processor would 
not give the correct result when re-interpreting the serialization. 
Perhaps the namespace could be serialized as "example:" and the local 
name as "mynamespace-myproperty", but is "example:" a valid namespace URI?

Is the bottom line that, in cases like these, a serialization must draw 
arbitrary lines in the string just so the name will serialize using a 
(namespaceURI, localName) pair?

Is there anything authoritative on this?

Cheers,

Garret