maybe we can 'memory map' some of it ? Subsequent locations in memory that you plot to a screen on over modificiation (b/w would be ok in this case I guess), a bunch of bitfields that get 'latched' whenever a key is pressed that the brainfuck mainloop has to walk by every now and then, resets and uses as branching points ?

That should do it.

It doesn't need blocking io, see. Polling latched inputs is the way the machines of old did it, it's a bit less efficient but it works.