Using the power meter in reflection I saw that the mode matching of IRMC was > 95%.

Flipping the INV/NON INV switch makes the IRMC lock to the mode with output power up to 1.01 mW (/1.68 mW input = 60%). So that works now. Actually, I did the same thing for SHG previously when it was showing similar behaviour...

Adjusting the gain on the IRMC servo shows very little change in the transfer function. For gain increase from 1.1 to 7, the unity gain frequency increases to a rather underwhelming 15 Hz. Now that I think about it, the IRMC PD (Thorlabs PDA05CF2) is specified for a -3dB bandwidth of 150 MHz (75 MHz linewidth). 

IRMC error signal has a DC offset of 150 mV for some (probably not good) reason (figure 1). Green mode cleaner error signal suddenly went to about 60% of what it was yesterday (figure 2).