Yeah, I use that page as a crash-test too :)
Anyway, I haven’t dived in the source yet but I’m very interested in how this is implemented. How much can you hook into the browser’s selector engine / how much functionality do you have to duplicate? (I took over the cssselect project for WeasyPrint, but it is in Python without dynamic updates so the context is very different.)