You should move focus from looking for a microcontroller to a microprocessor. You know, those +500Mhz ones.
Or, get some programmable logic. So you can apply hardware filters, something CAN and MAC (layer of ethernet) already have.
At what speeds does RS485 lose signal integrity? The distance should be reduced significantly with 20 MBaud, or not?
It needn't lose signal integrity, if it's planned right. Usually a bus is a double-end-terminated transmission line (remember coax thinnet?) with low impedance capable bus drivers distributed about (for a multi-master configuration). Aside from DC coupling and isolation, it's not really very different from 10BASE-T (20Mbaud Manchester encoding = 10Mbps), which is also doubly terminated (50 + 50 ohm to GND, 100 ohm differential), and reasonably strong (~1V?). You'll ultimately be limited due to HF and DC loss in the cable (unless you add boosters or repeaters).
wire capacitance and slew rate of line drivers are a key limitation.
Both edges and bits propagate just fine, when using compatible drivers (RS-422/485 outputs are not weedy CD4000 gates, or current-limited RS-232!) and suitable termination. The line can be much longer than the risetime, and for that matter, much longer than the bit time.
Speaking of 10BASE-T, the nice thing is it's self clocking. Most high speed protocols are synchronous, as mentioned. You aren't going to get away with that for two channels of RS-485 and a synchronous (e.g., SDO + SCK) protocol because no two cables will be well enough matched over a long distance that you can count on the clock and data remaining in sync. (If you're cheap and use unshielded CAT-5, you're even worse off, because the cable specifically has different velocity factors on the different pairs!)
Tim