Well, the CPU's internal flash would be mostly empty, while the I/O would be pretty busy with all that IDE, RS232, keyboard map, mouse readout, possibly a filesystem and drivers for god knows what else we attach to it.

This could be done in the way that main CPU queries I/O for OS and on failure falls back to uploading the built-in OS from its own flash.

(this way would simplify development a bit too - the system can bootstrap with just CPU and memory, without care about existence of any of the slave chips.)

Yes, we're mixing competences a bit, CPU steps into the I/O chip's field. But IMHO that's a pretty harmless move if we don't do that too liberately.