So it sounds like, the faster the target block time, the faster we need to retarget the difficulty.  I also like reward delay and graduate reward increase idea.

These parameters bounce of each other in interesting ways, but it doesn't seem  like a terribly complicated system to model.  The hard part would be finding the boundaries of the variables.  Things like minimum time required to request and start new work.  minimum latency requirements.  maximum data throughput to a casual users wallet.  That would give us limits on the model, which we could then fine tune to give the performance we want.

...

brain ran off the deep end here for a minute.  this may be completely unrelated, but i'll dump it here anyway, since we have people who actually think in this thread.  bare with me; i'm shooting from the hip here, and the coffee is only starting to kick in.

dynamic difficulty retarget seems a given.  what about dynamic rewards?  it sounds like a terrible idea...  i believe liquidcoin made an attempt at it, but hear me out a sec.  dynamic rewards plus dynamic difficulty would allow you to tune for a specific number of coins a day, regardless of network has rate.  that seems like a reasonable way to control inflation. 

... i had an idea here regarding arbitrage, but I just talked myself out of it.  either way, inflation is one of the critical variables that needs to stay well damped in a fast moving currency.