'm biased (being a Python expert but pretty rusty in Java) but I think the Python runtime of GAE is currently more advanced and better developed than the Java runtime -- the former has had one extra year to develop and mature, after all.
How things will proceed going forward is of course hard to predict -- demand is probably stronger on the Java side (especially since it's not just about Java, but other languages perched on top of the JVM too, so it's THE way to run e.g. PHP or Ruby code on App Engine); the Python App Engine team however does have the advantage of having on board Guido van Rossum, the inventor of Python and an amazingly strong engineer.
In terms of flexibility, the Java engine, as already mentioned, does offer the possibility of running JVM bytecode made by different languages, not just Java -- if you're in a multi-language shop that's a pretty large positive. Vice versa, if you loathe Javascript but must execute some code in the user's browser, Java's GWT (generating the Javascript for you from your Java-level coding) is far richer and more advanced than Python-side alternatives (in practice, if you choose Python, you'll be writing some JS yourself for this purpose, while if you choose Java GWT is a usable alternative if you loathe writing JS).
In terms of libraries it's pretty much a wash -- the JVM is restricted enough (no threads, no custom class loaders, no JNI, no relational DB) to hamper the simple reuse of existing Java libraries as much, or more, than existing Python libraries are similarly hampered by the similar restrictions on the Python runtime.
In terms of performance, I think it's a wash, though you should benchmark on tasks of your own -- don't rely on the performance of highly optimized JIT-based JVM implementations discounting their large startup times and memory footprints, because the app engine environment is very different (startup costs will be paid often, as instances of your app are started, stopped, moved to different hosts, etc, all trasparently to you -- such events are typically much cheaper with Python runtime environments than with JVMs).
The XPath/XSLT situation (to be euphemistic...) is not exactly perfect on either side, sigh, though I think it may be a tad less bad in the JVM (where, apparently, substantial subsets of Saxon can be made to run, with some care). I think it's worth opening issues on the Appengine Issues page with XPath and XSLT in their titles -- right now there are only issues asking for specific libraries, and that's myopic: I don't really care HOW a good XPath/XSLT is implemented, for Python and/or for Java, as long as I get to use it. (Specific libraries may ease migration of existing code, but that's less important than being able to perform such tasks as "rapidly apply XSLT transformation" in SOME way!-). I know I'd star such an issue if well phrased (especially in a language-independent way).
Last but not least: remember that you can have different version of your app (using the same datastore) some of which are implemented with the Python runtime, some with the Java runtime, and you can access versions that differ from the "default/active" one with explicit URLs. So you could have both Python and Java code (in different versions of your app) use and modify the same data store, granting you even more flexibility (though only one will have the "nice" URL such as foobar.appspot.com -- which is probably important only for access by interactive users on browsers, I imagine;-).
