Other Possibilities

Namespaces It wouldn't be difficult to create entities whose names were Qnames, relying on the declarations in the surrounding elements for their interpretation. It may not be what XHTML/SVG/MathML/etc. want to do, but this may make sense in organizations where there are lots of vocabularies with their own entity sets.
Context-sensitivity It wouldn't be hard either to define vocabulary sets which only applied in content held by an element of a particular namespace, and if you were ambitious this could use XPath.
Standardized sets In conjunction with or in place of any of this possibilities is the option of creating a standard set of character entities for Unicode characters which works across vocabularies. This feels unlikely to me but is certainly one way to attempt to resolve the character entity issue.
Further indirection Eliot Kimber's XIndirect work isn't about entities per se, but it suggests both the value of additional indirection and some XPointer applications in this space. (I won't implement anything further involving XPointer as specified, but you're of course welcome to do so.)
Bright possibilities yet undiscovered In the research for this paper, I had a sinking feeling that no one had looked in this dusty corner for years. I think there's a lot more worth examining here which I haven't found.