For lack of a great way to look at the low-frequency noise in the REFL PDs (due to the noisy DC outputs described earlier), I went back to the old idea of using a pickoff from near the laser output to look at intensity noise. The setup is a little bit hacky, since I just shot the beam through the side of the box with no viewport, but I think it's fine for this measurement. I took the light that was being dumped from the initial PBS. It was strong enough that even through the box I still had to attenuate it using an OD(=2) filter.
The spectrum I saw had increasing noise at lower frequencies, as should be expected I suppose. It didn't look quite right to explain what we see in the gyro signal, but I it wasn't obvious enough to rule out. I built a simple ISS servo by feeding it into the digital system, subtracting an offset, putting it through a 100-Hz LPF and feeding the output to the power adjust on the laser controller. I was able to get ~200-300x suppression below 100 Hz without the loop becoming unstable. The loop had essentially no effect on either the gyro noise or the TRANS_DC spectrum, leading me to conclude that the excess noise is not from input power fluctuations. I guess this is somewhat of a relief, as I have no idea how that would cause the noise we're seeing anyway.
Here is a before/after plot of three signals: the gyro noise, the TRANS_DC signal, and the ISS PD signal. NOTE: The ISS PD signal units are arbitrary, so the fact that the open-loop noise is right with the other traces around a few Hz is meaningless. As in the last few entries, I have multiplied the TRANS_DC spectrum by a phenomenological factor of 10 to show that it has precisely the same shape as the gyro noise at low frequencies (that is less convincing in this plot due to less averaging and higher FFT BW, but look to old posts to be made a believer).