<sigh>Well, that would certainly explain why comparing the results before and after didn't warn me that something had changed...</sigh>
Made some progress on the LRE10/LRE21 bug anyway... The namespace in question is an Extension Namespace, and so is flagged as being "excluded". But we have an attribute which uses it, so we should override that and declare it. That override is apparently supposed to be handled by the last-minute call to ensureNamespaceIsDeclared(), issued when the attribute is written out. That call _is_ adding the namespace declaration to the stack, but is not causing it to be written to the output document. We need to either have it do so, or have it pass back a hint that doing so would be appropriate. (Our namespace architecture is just a bit baroque. I think that's partly because it's written in terms of prefixes-mapped-to-namespaces rather than converting to namespace URIs immediately and operating primarily in that space thereafter. We might want to reconsider that design at some point.)
