Summary:

1. While noisier, I was able to control the arm lengths to ~30pm RMS(!) using the green ALS beats as error signals (cf. ~10 pm RMS with the IR ALS system).
2. The PRMI could be locked with a CARM offset applied.
3. When lowering the CARM offset, I saw an increase in the in-loop ALS error signal, just as I had with the IR beat.
4. IR TRX / TRY unsurprisingly did not stabilize in any meaningful way.
5. The noise increase seems to have some periodicity along the frequency axis - need to think about what this means.
6. Since there is no apparent benefit to using the green ALS beats, I restored the IR system. The green PDs should still retain somewhat good alignment if one wishes to do a comparison measurement.
7. While the shadow sensors of the ITMs report elevated noise, it is unlikely to be responsible for the cavity moving by the amount suggested by the elevated ALS error signals because of the digital low-pass filtering and 1/f^2 of the pendulum.
8. I confirmed that the ITM shadow sensors do not report elevated noise when the PRMI is locked such that the carrier is resonant. In this config, there is comparable circulating power in the PRC as to when the CARM offset is reduced to ~0.
9. The fact that the IR and green beats both show similar increase in noise suggestes that the cavity length / laser frequency is in fact being modulated, but I still don't know what the exact mechanism is.

was worth a shot i guess.

Trawling through some elogs, I see that this kind of feature showing up in the ALS CARM is not a new problem, see for example here. But I can't find out what the resolution was.