I noticed the update to wiring.c in your latest v20, but haven't had chance to play around with it yet. Will try and find some time soon.
I've also been meaning to have a look at using CLKPR as a runtime reference instead of F_CPU during compilation. Would be nice to switch speeds on-the-fly and still retain some timing functionality, at least in ms.
With a 256 prescaler (37.5kHz) each clock cycle is 26.66us, so an ASM loop using 38 clocks would return after 1013.33us, which is close enough given the chip tolerances. Repeating the loop 256/CLKPR times would give the same 1ms regardless of operating frequency. (9 to choose from and no fuses to worry about!)
Seems fairly simple if using the internal 9.6MHz oscillator, and not much more complex with an external clock. (Adjust for that with F_CPU at compile time.)
Downside is that it's not going to work with us delays, but millis() and micros() might be possible with some work. The upside is that it would offer a lot of different PWM frequencies (20 I think) from 37.5kHz right down to 0.07Hz.