the dsPIC's I'm looking at are FAR more overkill than any 8 bit AVR
A PIC18 is also an option, but this was also going to be a core unit for a larger motor controller in the future. I'd probably also roll a model based on the PIC18F2480 (~$10 Digikey vs ~$15 for the dsPIC), depending on pin count requirements. If I do a lot of the code in C, it should be fairly portable between architectures (although reinforcing the need for my overkill in processor selection).
The motor control unit I was planning was a dsPIC30F4011, QFN package. I might be able to go to a 30F4012 (same package, a bit cheaper) since I may not need I2C anymore (duplexed with the CAN transciever on the 4012, separate on the 4011). Hardware H bridge PWM and hardware quadrature make a lot of the processing overhead bare. Which is nice, as I can then implement a bunch of additional modes I want to (constant torque, low jerk, synchronous feedrate, transparent multi-drive control, etc).
And yes, a CAN chip at 6x6mm of board space is large, with an 8x8 44QFN next to it (or a 6x6 28QFN). I still think it's possible, it'll just take some work.