any other metadata has to be referenced from the representation of the resourceThis is good. Semantic nets typically connect data (e.g. "frames" and "slots") and metadata (e.g. "facets" of a slot) directly. I would expect this to be so for a semantic web as well.
use of REST is not always appropriate (even if theoretically it can be used almost everywhere)This is almost certainly so. But I suppose it is more broadly appropriate than its current range. And less clear is what is more appropriate in those other cases. Certainly there are cases where pre-existing systems are more appropriate. But what of the up-and-coming distributed systems technologies?
REST might be inappropriate especially if the integrated legacy system defines its own non-RESTful abstractions (e.g. SAP's BAPI).Certainly there will be these cases. But I'm not sure SAP R3 can be put in that category. First, SAP has provided at least a couple of web-services-like interfaces to R3 and newer components. This is settling into the Netweaver/XI integration platform.
I don't know much about this yet, but I don't see why REST would be incompatible. One of the best presentations at OSCON 2003 was convincingly about a Perl and REST interface to R3.
use of REST is especially appropriate for generic system-level services that are expected to be used by robots (registry, security, management, reporting)I think you are underestimating what a "robot" (i.e. an "interpreter" of semi-structured information) can do. See the Experiments with OVAL paper for a variety of examples using very modest software.
REST is appropriate for hypermedia, of course ;-)Maybe a good discussion could be about what is suitable for representation as a hypermedia?
building REST applications, WS-* should be used wherever possible to ensure interoperabilityMaybe this should be turned around a bit to say that WS-* should be defined to be compatible with the web.
RDF/OWL is more important for REST applications than it seems todayI think generally we need to learn how to build more evolvable application-specific semantic web interpreters. I know a little about RDF and less about OWL, although I am familiar with older semantic net languages. I think this statement could be so. But again referring to OVAL, a lot can be accomplished without getting to deep into heuristics.