I don't think an XSLT processor will ever be able to determine
statically that the output of a stylesheet is guaranteed be valid XHTML.
There may well be cases where it is able to determine statically that it
won't be valid, however.
Any specific reason for that (first) conclusion - for instance, do you
have any feedback from XQuery implementors in this area? Moving even
further off-topic, any feedback on how suitable Haskell is for XML
processing? It might be interesting for readers of this list to have
guidelines for what problem areas XSLT 1.0 / XSLT 2.0 / SAX / Java (or
C#) / Perl / Haskell / (whatever) are best suited for.