I have to say that
I can't imagine why people insist in doing this, they almost invariably
shoot themself in the foot at first.
Actually I think it can be seductive for one thing, which is to use xslts model
for progressing through an xml document in situations where one has to build
interpreters for the a document, stuff like CIM operations over http. An
argument against it would be to output javascript and eval() that but I am not
sure about how that compares with internal evaluation. for example
Cim over http -> url -> call xslt -> eval output -> build cim response -> send
response
CIM over http -> url -> call xslt -> eval parts of tree, build response -> send
response.
At the same time it seems like a seductive model it also seems totally
wrong-headed, but then again I have had successes in various instances following
the wrong-headed path.