I have made the test of decreasing the SSFS gain during LN3 to see which contributions to the sensitivity come from the frequency noise. The results are shown in FIgure 1, the blue curve corresponds to standard conditions (10kHz) and the violet one to reduced gain by a factor 4 (2.6kHz).
The GPS are: 1211808164 (blue, 2min), 1211808437 (violet 2min).
The worrying thing is that above 2kHz the sensitivity is coherent with frequency noise, and above 4.5 we improve the sensitivity when the SSFS loop is lower, which points to a problem in the SSFS loop.