We noticed that the channels readout with shaping filters (2-12 Hz, Pr_B1_d2_40K) show more coherence: see left plot of 1st figure. We can therefore suspect electronic noise.
On the first plot the purple and green curves show the expected noise from ADC non-linear noise for the typical observed low frequency component of B1_ACp (purple=upper limit, green=lower limit). This could explain the loss of coherence.
The 2nd figure shows again the expected up-conversion noise (purple and green) while the black is the difference between d2 and d3 photodiodes, measuring the noise not coherent between photodiodes like shot noise or electronic noise. The incoherent noise is indeed consistent with the expected ADC non-linear noise.
Although the gain in sensitivity won't be as one would naively expect from these curves (due to the effect of the control loop) it would be better to use channels with shaping filters for the d2 and d3 readout.
Therefore I would say that this ADC noise might be important at intermediate frequencies 20-60 Hz.
. purple: line@379Hz with an amplitude of 5e-3V
. green: line@379Hz with an amplitude of 1e-3V