Figure 1. The glitch is the same in the Audio and DC channels, but it is different in the blended channel. Looking at the transfer function between the blended channels, it is probable it has the same shape as L/(1-L) or something similar, where L is the DARM loop open transfer function. This would mean that the glitch appears in the blending computation in one of the two photodiodes.

The blending filter design looks quite different compared to the old photodiodes, with a pole at 0.006Hz for a channel sampled at 20kHz, the large ratio between the two frequencies could be an issue causing numerical problems.