Just to confirm that my noise estimates make sense, here's a plot of the notnormalized QPD signal that gives the X motion (sum and difference of all four quadrants):
This is the signal after compensating for the whitening filter. If I remove this compensation, the following plot gives the noises in terms of the voltage directly in input to the ADC (or in output of the analog board):
So the total "dark" electronic noise is about 13 uV/rHz.
I did a roughly estimate of the sources of electronic noise:
 QPD dark current noise, from datasheet, at the peak sensitivity is equivalent to 2e14 W/rHz, or 2 nV/rHz at the output of the TI stage
 First stage: JohnsonNyquist noise of the TI resistor: 58 nV/rHz
 First stage: output voltage noise of the LT1124: 3 nV/rHz
 First stage: input curent noise of the LT1124, converted to the output: 60 nV/rHz
So the total noise at the outoput of the first stage is about 84 nV/rHz. The second stage adds a gain of 30 at high frequency, and negligible noise. So at the output of the whitening we have 2.5 uV/rHz. The DRV135 adds another gain of 2 and a neglegible output noise.
So the total electronic noise at the output of each quadrant is 5 uV/rHz. Since we are combining four of them, the total expected electronic noise is 10 uV/rHz, which is not too far from the measured value.
We are basically dominated equally by the JohnsonNyquist noise of the TI resistor and by the input current noise of the LT1124. No gain to be obtained by changing the whitening.
Quote: 
I measured the noise sources limiting the QPD sensitivity. Unfortunately, I had to do some MATLAB tricks to get rid of the glitches: basically I load the data directly from the raw frames (NDS access to data is not working yet) and remove all jumps in the signals that happen in one single sample and are larger than a manually tuned threshold. This is not perfect, but it's enough to give us a rough idea of the spectrum of the QPD signal. The following plot shows the QPD_X signal (in units of disk motion, radians) in a few situations:
 Blue: normal (laser on, room lights on)
 Orange: laser on, room lights off (including the vacuum gauge)
 Yellow: laser and room light off
 Purple: same as above, but I switched off the PC monitors too
 Green: QPD electronics off (this is the ADC noise)
The total power on the QPD is 30 uW, which correspond to a shot noise limited sensitivity of 4.3e12 W/rHz. Considering that the signal is the quadrant asymmetry normalized by the total power, the shot noise limited sensitivity is sqrt(2) * SN / Power which once calibrated corresponds to 1.1e10 rad/rHz.
The following plot shows that shot noise is the dominant source, followed closely by the electronics dark noise. The total agrees perfectly with the measured background noise above 2 kHz. Below that we have some leakage due to the large turbopump peak: this is due to FFT limitations but mostly to unsuppressed glitches.
From the QPD datasheet (Hamamatsu S5981) I learn that the noise equivalent power should be of the order of 2e14 W/rHz at the sensitivity peak, so probably a factor of two or so worse at the HeNe frequency. It's still much lower than the measured dark noise.
This sensitivity is already pretty good, but we can improve it by increasing the power on the diode. Indeed, 30 uW corresponds to about 2.7 V after the transimpedance, so we could increase the power by a factor 4 and win a factor 2 in the shot noise to dark noise ratio. Probably not worth it, since it will give us only a 30% gain in high frequency noise.

