We noticed that increasing the RF9 modulation index increases the noise in DARM. What's more, the shape of the excess noise very closely follows our mystery noise (figure 3), and indeed, just as the mystery noise, the excess noise is coherent with REFL RIN and POP RIN in much the same way as our mystery noise (except stronger, figure 1). 

Next we injected 1e-6/rtHz of RF9 RIN at the modulation depth stabilizer. Using both the injected RF9 RIN and the self-generated noise, we measured the transfer function to DARM. They agreed perfectly (figure 7):

REFL: 8e-12 m/RIN
POP:  5e-12 m/RIN
MOD:  9e-14 m/RIN

Figures 3, 4, 5, and 6 show the noise for 3dB increased RF9 modulation depth (red), the injected RFAM (blue), the reference noise (yellow) and the DCPD cross-correlation estimate back-projected the the diode. Figure 3 is DARM, figure 4 is POP RIN, figure 5 is REFL RIN and fig 6 is at the modulation stabilization box (this one does not show the self-generated RFAM noise).