Re: glossary sorting/indexing question

[...] Currently I use xsltproc for a large number of small transforms
because the performance edge over a 2.0 processor adds up to a
significant time saving in the full build [...]

In case you actually use the command line processor instead of the
LibXSLT API, additional time saving can be made using one of programming
APIs, like Perl or Python.

Same story, of course, for Saxon. Saxon is fast - it's just the JVM
startup overhead that slows things down when incurred repeatedly via a
batch script. So if you program your large number of small transforms
into the Saxon API, I can't imagine it'll be that much slower than
LibXSLT; depending on the nature of your transform, it may even be faster.