As some experimental code is now written using stax, I continue to believe that the result will be less robust than the equivalent metadata-for-xb code. 450ms of startup time [the exact amount is not known yet obviously] is not worth a decrease in quality.
Looking at the code, it would be possible to make the new classes more or less equivalent to the current metadata classes, which would allow reusing the code for augment/merge between fragments and annotations that is needed for Servlet 3. So that's a good, and the sticky point is now mostly with the hard core XML parsing.