Either qualified names in attribute values have to be banished, or some new version of XML should be created. Now not much people want the latter, so. Having qualified names in attribute values eliminates namespace prefix rewriting. It also requires you to know in advance which attributes can contain qualified names. Possibly by looking it up in some schema. Of course, I’m not the first to find out.
At one point the Atom WG faced this problem. The resolution can be found in section 22.214.171.124 of RFC 4287. Indeed, “strings” or IRIs. IRIs are for unregistered extensions, the other allowed values can be found in the Link Relations Registry.