So: A a TeX compiler that can be called from JS can only be an intermediate solution.
I think http://txtjs.com/ goes in the right direction.
I don't really think calling TeX from js is a very good idea, nor do I really think canvas is a good idea either.
I do wish Adobe had gained more traction with css regions[cr] -- although I also understand some of the arguments against them[ch].
I still think the idea is good: css for style, semantic html for ... semantics -- and "layout html" for layout. I think pairing semantic markup with css columns is just a bad idea -- and it gives the kind of "half-power" that initially lead people to use tables for layout.
So Lie is right in that css regions aren't a great fit for html[ch] -- but I still think they're the best fit for html I've come across. I might be convinced that css regions are the kind of things that fit well with js/poly-fills[pf] -- although I'm sceptical. We know fonts are turing complete and have had security issues -- I don't see why we should need to implement core layout with js. At least I suppose "rogue" software vendors can choose to support css regions[cr] -- For eg an e-book reader based off of a subset of html+css, css regions might be a good fit, while avoiding the complexity of adding js support, and trying to make that secure (including proof against denial of service etc).
Thanks for the link to texjs, though -- I'll definitely have a closer look.