We only produce an abstract for articles in the main namespace that are not redirects and that have a content model of text or wikitext. For Wikidata, all items in the main namespace are Q-items with content model wikibase-item.
If this is not something we expect to change, then there' s no point in spending 36 hours generating a bunch of files that contain for each entry, <abstract not-applicable="" />. We should just generate an empty abstract file and be done with it. I guess we'd generate the 27 empty partial files and one complete 'recombined' file, each containing only the mediawiki and siteinfo information and the mediawiki footer.
I'd like to get the input of folks on the xmldatadumps mailing list as well as @hoo to see what people think.