Today I checked whether REFL SERVO ERR is a trustworthy readback by comparing it to an analog spectrum from OUT2 while the IFO was in a single-sensor state (so that OUT2 == REFL SERVO ERR).

What matters is OUT2, since the CARM loop is fully analog for fast feedback.  For our projections, we use REFL SERVO ERR.  I was wondering if that is a valid thing to do.
Below is a comparison spectrum of the two.  Seems fine.

OUT2 is worse at low frequencies, this is likely because the SR785 noise floor was too high during the time of the meas.
REFL SERVO ERR has some peaks at 2 to 3 kHz.  Robert pointed out these are probably from aliasing down from the sampling frequency (16 kHz) and this huge peak forest at 14 kHz.  
In any case, they are not present in the actual loop.

Frequency noise projections using REFL SERVO ERR should be valid for all frequencies.