I've now figured out how to create an extension module for eXist, following the instructions here. These are some things I've learned:

• The only practical way to do this is to work with your module code in the context of the eXist tree, in $EXIST_HOME/extensions/modules/src/org/exist/xquery/modules.
• You can use a non-eXist namespace -- I'm using http://hcmc.uvic.ca/ns/usm -- but it seems safest to use the eXist package structure, so my package is in org.exist.xquery.modules.unisimmetric.
• All the extension modules are built together into a single jar called exist-modules.jar. You can build this jar alone, using build.sh extension-modules, then drop that jar into an existing eXist instance (although if the new jar was built with a substantially different version from the rest of the code, there could well be problems).
• To turn on your module, you add a line to the conf.xml file like this:

  <module uri="http://hcmc.uvic.ca/ns/usm"                        class="org.exist.xquery.modules.unisimmetric.UniSimMetricModule" />
  

  along with the other modules.

I'm not yet happy with my module, and I'm still working on it. In particular, I'm not happy with the scores it's generating, and I think this might be something to do with other bits that get included in the GZIP stream, such as a header; if I can figure out how big those are, I can remove them from the calculation. The highest difference I seem to get is around 0.53 with completely dissimilar strings, so it seems as though the results are being compressed into a range much smaller than 0-1.