I find that public support forums like IRC are useful even when nobody answers your question -- even when nobody even acknowledges that you asked a question. This is because that is what often happens (at least, to me) when the question itself is totally wrongheaded. Getting the online equivalent of crickets chirping in a silent auditorium tells me that I need to go back and check the assumptions that led me to ask in the first place.
Case in point, I've been trying to get someone in Freenode #xml to explain how entity dereferencing could possibly be sanely implemented in the context of the DOM. Because an entity reference can be expanded, as per the DOM recommendation, into practically any of the full set of DOM Node types, I wanted to know what a reasonable DOM implementation would be expected to do with a situation like:
<!ENTITY ent ""><foo attr="baz"></foo>">If the expansion lead to:
...
<foo attr="bar&ent"></foo>
<foo attr="bar"><foo attr="baz"></foo></foo>...that would be problematic for my DOM implementation, since the parent of the entity reference and any nodes that came after it in the document order could have the meanings modified or positions changed, and I'd probably have to pass a serialized version of the entire document back through the XML parser. The proposition filled me dread.
But when I didn't have any luck on IRC, I went back to the DOM recommendation and eventually realized that I didn't have much of a problem after all -- regardless of where entity references can appear during XML parsing, when you're manipulating the DOM they can only be added as proper children of nodes, and not as children of text or CDATA sections. The only case that has to be taken care of is something like this:
<foo attr="bar">&ent</foo>No sweat!