Important detail: only in combination with an arbitrarily large HTML file, and only with arbitrarily many user interactions (i.e. the user needs to manually increment the state).
As such it's a pretty meaningless type of turing completeness. So, whereas in general it's undecidable whether a program completes (the Halting problem) if it's written in a turing complete language, for the case of HTML+CSS it may well be decidable for several reasons. Firstly; because the HTML file will be of limited (and in practice small) size, and secondly, because the interesting question is whether you can reason about the state _now_; not whether the document might reach arbitrarily many states were the user to perform infinitely many interactions.
Turing completeness makes hard to reason about a program; and indeed the weaker a language the easier it can be to reason about it - or, more practically - the easier it can be for the browser to determine boundary conditions and to simplify interpretation of the page. Indeed, ironically, Tim Berners-Lee points out the advantages of using weak languages: http://en.wikipedia.org/wiki/Rule_of_least_power
Saying that HTML+CSS is turing complete is kindof like saying finite state automatons are because the number of states is unbounded. That's entirely beside the point however; the point is that I can evaluate a _given_ FSA without turing completeness, not that an arbitrarily large FSA might take arbitrarily many execution resources (I mean, well duh).
Any given HTML+CSS combo is thus bounded in two ways; both by its own size, and by the fact that I don't care about eventual possible states reachable after infinite user interaction but just about the current state (or states very close to the current state).
http://news.ycombinator.com/item?id=2300836
https://github.com/elitheeli/oddities/blob/master/rule110-gr...
in combination with HTML at least...
(Although as others have stated this game is really just a state machine)