Working on the Java classes alongside the XQuery code that's calling them, I've now extensively revised both, to give results which seem to work well, and give valid, usable output. The list of phrases to highlight is now provided as an array by the Java class, and the XQuery iterates through that array to create a list of items, resulting in actual XML elements rather than textual data. When there's no search string, a dummy text body is returned for each document header, so that the resulting corpus file is valid.