This person is confused. He's citing a Ted Nelson paper about separating these things into layers (content, structure, & special effects), while personally advocating that we mash it all into unicode.
(Cf. how the cross-reference stream in PDF files makes it painful to edit objects in them, even when the files are nominally encoded in plaintext.)
He then goes into how a separate styling layer can assist with transcluding text from other people's work while modifying the style. But style variations are hardly the only legitimate changes typically made to direct quotations: people often want to modify capitalization or punctuation, elide portions, or insert bracketed notes. And at that point, you're modifying the content as well as the styling, so style-only modifications would be very limiting for that use case.
As for the structure layer, this would have the same issues as every other attempt in the last three decades to create a semantic web or whatever. Authors don't want to spend their time carefully curating metadata that 99.9% of readers won't care about, while bad actors want to game their relevancy metrics through any mechanism available.