For hardcore users I don't think this tactic suggestion thing is a good idea, for example, how does it deal with custom ltac tactics (cf Chlipala's bedrock)? To prove me wrong one could test the idea on, say, compcert's development and compute how often the next tactic is among the suggested ones.

On the other hand, One common problem in large proofs is having too many hypothesis in stock, one super nice extension would be to quantify the relevance of each and color/display them accordingly, leaving the option to move the 'tolerance' threshold for display. This relevance metric would have to be aware if lemmas available (of A -> B is proved by a lemma, and B is the goal, A is relevant).

My 2 cts.