I like the idea of flat address spaces with memory protection only. There are a variety of ARM processors that are set up that way, like the ARM940T which has 16 regions that can be configured in size, position, priority, etc. Using a large background region with several smaller high-priority regions that overlap it, you can run several tasks with enforced private memory and shared memory spaces. That way you get some of the benefits of an MMU, but with fast context switches and deterministic timing.