There's actually been several HTML UI projects that "break compatibility with the web" in varying degrees, with some breaking almost completely (eg, Sciter and libRocket) for the sake of simplicity and performance.
The goal of Ultralight was to see if we could strike a balance between modern HTML5 compatibility (eg, actually being able to render the Alexa Top 50) and reducing size and complexity for use in an embedded use-case.
I think we've struck a good balance so far but can go much farther. We plan to support much more of the web platform while still remaining small by use of shared modules (eg, for WebAudio/Video) so users can pick and choose how big of a distributable they want based on feature support.
As an aside, I took 2 years off just to explore if this was possible, partly funded by my previous work on Awesomium. This project is the result of that research and development.