I agree that markup/stylesheets is what most strongly needs to be improved. I've only just started experimenting with it, but Biber, positioned as a BibTeX replacement, does really seem to be an improvement for the citation piece of the puzzle; its stylesheets look like something a mortal might be able to edit. It can also be used to generate HTML output, which has long been pretty crufty with BibTeX (I already have paper lists in BibTeX, so I should be able to generate things like my web page's publication list, and be able to customize the generated HTML, maybe even have some sorting/grouping options, all of which Biber can do).
If I had to predict, I would guess that something like HTML+CSS, or possibly a Markdown-ish input language with some kind of CSS-ish stylesheet, will eventually overtake TeX in significant areas, once the PDF renderers get good enough. Maybe someone will even find a way to pipe it into TeX as a renderer, as you suggest. A plus of that workflow is that it also makes it easy to produce good-looking HTML versions of articles, which is getting more important. And the "math in HTML" question is (finally) converging on a constellation of acceptable solutions. Advanced typography in HTML is slowly creeping forward as well.
If I had to predict, I would guess that something like HTML+CSS, or possibly a Markdown-ish input language with some kind of CSS-ish stylesheet, will eventually overtake TeX in significant areas, once the PDF renderers get good enough. Maybe someone will even find a way to pipe it into TeX as a renderer, as you suggest. A plus of that workflow is that it also makes it easy to produce good-looking HTML versions of articles, which is getting more important. And the "math in HTML" question is (finally) converging on a constellation of acceptable solutions. Advanced typography in HTML is slowly creeping forward as well.