It's taken a lot of trial and error, but I've found a path through MATLAB loops that seems like it may be stable at all points.

CAVEAT: This doesn't give any indication as to why we weren't able to turn up the AO gain more last night, as far as I can tell, so it's not all good news. 

However, it's still ok to at least have a plan that works in simulation... 

Based on the location of the optical resonance peak in the CARM plant, we estimate our CARM offset to be 200pm. I haven't simulated TFs there exactly, but do have 100pm and 300pm TFs. This procedure works in MATLAB starting at either, though 100pm is a little nicer than 300. MATLAB data and code is attached in a zip. 

The steps below correspond to the attached figures: Bode plots and step response of the Loop at each step. 

0. [Not Plotted] DCTrans sensing, MCL actuation on CARM. FMs1,2,3,5,8; UGF = 120. (DARM not considered at all)

1. AO path just turned on. Crossover with MCL path ~ 3.5kHz. 
2. AO gain increased. Crossover ~ 500 Hz. There are now multiple UGFs! Handling all of these in a stable manner is tricky. 
3. AO gain increased. Crossover = 150Hz. [No simulations with a higher crossover survived the next steps]
4. Compensation filter applied to MCL path; 1 real Zero at 105Hz and a pole at 1k. From a TF point of view, this is sort of like switching to REFLDC, but the SNR at low frequencies is probably better in TR signals at this point. 
5. CARM offset reduced to 30pm. (This smoothens out the optical plant resonance.)
6. Overall gain increased by factor of 3. There is now just one UGF at a few kHz, above the optical resonance. From here, gain can be further increased, boosts can go on, offset can go way down. In reality, we should switch to a single error signal once we're back to one UGF, and go from there. 

#4 Seems like the most sticky part. While both sides of this look stable as far as I can tell. I feel that flipping from the red phase curve to the teal might not actually be ok, since they are on either side of the bad phase of 0 degrees. It isn't immediately evident to me how to easily model the transitions between steps, rather than just the stability of of each step in the steady state.