My nine test files now convert without errors (except the missing reference ids that refer to the bibliography items in the reference list, which are not converting yet because they're back matter). So two-thirds of the job is done. There are some oddities in the model structure of NLM -- for instance, every section (<sec>) must have a <title>, which seems ridiculous, and links (<xref>, uri and <ext-link>) cannot contain abbreviation tags, which seems pointlessly restrictive, when they can contain bold, italics etc. However, that's not really my problem, except that it requires me to throw away some information during the conversion.