Tweaked the XSLT for actual use (adding utility templates to preserve whitespace between PIs etc.), and SK has now tested this with real data. She has some suggestions for improvements, one of which might be achievable, although it will require a structural rewrite, but results are good so far, and most cases of deviation from desired results are in situations where there is no mechanically-detectable context which could be used to change the outcome. Can this now be extended to pron/segs, where there's even less context because hyphenation is not there?