I was imagining more like some kind of primitive CPU mode that sends every single RAM request to a software exception. Something you would normally have turned off but available for truly heavy duty debugging.

That sounds pretty similar to what Valgrind does, basically runs your program in a VM with that kind of instrumentation. The details of how it works are way past the boundaries of my own knowledge though :)

The operating system could effect this by not providing any page tables to the CPU, so every memory reference faults and the OS has to handle them all.