I decided to generate not just the linkGrp structures but a fully-expanded list of lists of variants, each with a link to the actual document from which it comes, constructed from its title, as the body of the variant spelling output. This saves additional steps in downstream processing. That's now working and valid.