You can already do that with NaCl - or even asm.js with eval, I bet it wouldn't be that much slower. But current JITs apparently gain a significant amount of performance with inline caching, which expects code to be able to be rewritten very quickly.
Edit to respond to your edit: although it would be cool to be able to have a "sandboxed mode" that can somehow be switched out of more cheaply than an interrupt, the whole thing seems like a massive hack to me. After all, NaCl does not take advantage of its pseudo-ability to do so: NaCl code runs in its own process and already incurs a context switch whenever it communicates with the browser, so there is no inherent hardware reason NaCl couldn't just run directly under the kernel and have the kernel provide the same level of sandboxing as the NaCl runtime currently does. It's just an issue of getting such an approach to work portably with existing kernels... hardware support might be able to make it easier to get that to work, but it's probably unnecessary.
Edit to respond to your edit: although it would be cool to be able to have a "sandboxed mode" that can somehow be switched out of more cheaply than an interrupt, the whole thing seems like a massive hack to me. After all, NaCl does not take advantage of its pseudo-ability to do so: NaCl code runs in its own process and already incurs a context switch whenever it communicates with the browser, so there is no inherent hardware reason NaCl couldn't just run directly under the kernel and have the kernel provide the same level of sandboxing as the NaCl runtime currently does. It's just an issue of getting such an approach to work portably with existing kernels... hardware support might be able to make it easier to get that to work, but it's probably unnecessary.