I couldn't even support such an idea in principle. For me, one of the cornerstones of XML's value is in completely determining the lexical space. I think leaving entity interpretation to applications would kick a huge hole in that.

Actually, it strikes me as quite an intelligent layering. If xinclude processing can be done in a different layer from XML parsing, why not entity expansion? In fact, it's strongly arguable that it SHOULD be a separate process; one of the use cases for entity expansion is to give a level of indirection so that the same entity reference can be replaced by different text at different places/times, and that works much better if it isn't embedded in the parser.