Your proposal is interesting, and I agree the confirmation time is definitely a point of some friction. I like litecoin's 2.5 minute target, but even that could get unrealistically large for things like high frequency trading. Definitely worth considering. One of the potential weaknesses is it creates an incentive to oscillate the difficulty. This incentive already exists, but you need a significant amount of hashing power to benefit from it in the current system. Bounding the weight deviation seems an effective way to keep that incentive down.