It is possible to detect correlations between questions (or even individual steps) and to implicitly determine skills at an even more fine grained level than that - given sufficient data. We aren't aiming to get that specific right at the moment - indeed, we mightn't have collected enough data yet to do that - but there's no reason why we shouldn't be able to do that

Having tagged data is extremely useful for designing these kinds of algorithms, but we would only require a proportion of the data to be tagged. So it wouldn't be anywhere near as costly as you expect