Actually... you do not need that fast a network. You can converge state incrementally rather than transfer it in one go. Xen and VMware do this. Surely the cost of a fast SAN and interconnect would quickly outpace the (imho minor) savings from smaller memory allocations.
I'm sorry but statements like that freak me out a bit.
Do I get a warning when my instance is going to be migrated instead of resized, and an estimate of the downtime?
Because unless you found a way to defy physics you are pretty surely not migrating e.g. a 4GB Ram/40GB disk instance in anywhere close to 2 seconds.