The VML APIs that've been in IE forever are all pretty damn fast individually, but displaying the results slows the rendering down significantly. It's like how straightforward PHP is damn fast as a templating system, but absurdly slow for implementing a high-level templating system in.

Besides, text selection wouldn't work.

Have y'all managed to avoid implementing a GDocs-style system for hiding text underneath bitmaps for selection?

Rendering an HTML element to an image isn't inherently slow; Safari does the same to hardware accelerate CSS animations via Core Animation and it's quite fast there. Of course, it's entirely possible IE's implementation is terrible.

I'm curious about the other issues you had as well.