1. HTML is designed for representing webpages, not books. Your intermediate representation will become HTML's DOM, which is likely not well suited to the book use case. You'll either need to use an existing HTML-to-DOM parser or build your own. One big limitation of the DOM, as I see it, is that it doesn't have the concept of a page number.
2. HTML is not extensible in the way the author talks about. This is the "semantic linking" thing I meant. In rST, you can define entirely new top-level constructs: he defines an exercise which has a solution and will be a part of a solution list. The semantic linking part is that other parts of the document will refer back to these solutions. Some output formats will have all solutions inline, with the exercises, while other output formats will collect all solutions in one place at the end. HTML is not flexible in that manner.
What is extensible is XML. It's entirely feasible for XML to be the input, but for better-or-worse, most people are even more turned off by XML than they are by something like rST.