Just guessing: (1) it's not actually "data" and this provides more meaningful cues, (2) it provides a bit of namespacing so as to not step on the end developer's own data-attributes, (3) turns out that it doesn't actually break anywhere after all, as proven by many JS frameworks at this point. I tried briefly to track some standards down for you, but I know this gets discussed at length on the Angular mailing lists if you really care for the nitty gritty.
