You use the MMU to create virtual address spaces and the kernel to provide a message passing interface. The MMU may cause you to incur some latency but in exchange you get some pretty good isolation between components of your system. Even more so if your kernel is proven using formal methods.
There's no need to get snarky. I've used similar parts before, for example TI's Hercules RM4 does not have an MMU either and so what I said wouldn't apply there. But that doesn't mean you would never have an MMU.
Indeed, my apologies. I actually quite like the idea. It reminds me a lot of Erlang, where you set up a bunch of small processes and if any of them die, you can a) try to recover quickly, or b) run in reduced functionality mode. I just haven't been fortunate enough to work with embedded processors where it'd be a good fit; they've either been on the small side (Cortex M, MSP430, AVR) or on the big side (iMX6, AM335x) and just run full-blown Linux.