I ran one test with an arbitrary load setpoint and the speed set to 1400rpm. This was running on only 200v so I was probably right up against field weakening at this point. Below is a quick plot of the results.

The blue line is the Iq value driven by the speed control loop of the motor acting at the load. As the torque of the driving motor increases, the Iq of the load must increase as well in order to maintain the speed setpoint.
The red line is the offset angle of the opposing (driving) motor.

Instead of using Iq and Id decoupled, Id always has a setpoint of zero and Iq is the only current setpoint used. The offset angle is the value (4095 = 360 degrees) which the dq axis is offset from being aligned with the motor angle (a-axis).

For the example plot above, the sweet spot falls at about -50 (only 4 degrees or so). A small number which makes sense for the low speed and low current conditions (<15A).

Note that the driving motor was actually set to negative torque (trying to slow the load down). This is why the offset of the driving motor is negative and the Iq current of the load is positive (representing positive torque). This condition simulates regenerative braking.

My experiments will extract the offset angle appropriate for a given Iq and speed. This will go in to a 2D look-up table. So all I have to do to control torque is convert the throttle input to an Iq setpoint, then that pull the associated offset angle from the LUT. This eliminates a whole set of trig operations that would need to be done in the microcontroller if I tried to control the angle by breaking it in to an Id vs Iq relationship.

If the relationship ends up being linear (or very close to it), I could just use a simple scaling and offset equation with no divides. Even better than a LUT, but I'm doubtful it'll be that simple.

This experiment is intended to get the confidence I need that it's all going to work before committing to tearing in to my car and doing the conversion. It should give me enough to make it driveable and give me most of the performance it'll be capable of. I might fine-tune it on a vehicle dyno someday depending on what operating points I can hit with the transmission alone. I'll be limited to the capability of MG1 which is somewhere around half of MG2. So I won't be able to fully characterize MG2 with this method, but hopefully I can extrapolate based on MG1's results. Time will tell.