Remeasured the frequency noise with the 785 after the visibility improvement. Assuming my calibration to Hz/sqrt(hz) is correct, the measurement is still yielding several orders of magnitude larger than what rio claims it to be. Investigating this today.
For reference, the SR785 measures a quantity is V/sqrt(Hz). To convert to frequency noise I do S / (L/c * B*2*pi) where S is the measurement, L is the path length difference between the two arms and B is the fringe amplitude. V/sqrt(Hz) / (m/(m/s) * V) = Hz/sqrt(Hz).
The right axis (Hz) on this plot has a different frequency dependence than the left axis (Rad). So be careful with the specification data. I agree otherwise that there seems to be an excess. It might be intensity noise that needs to be canceled.