We understood at least one problem that might have caused problems: Since the nominal frequency of LNFS differs by a few Hz from the demodulation frequency generated by the look-up tables in the demodulation boards, the phase is de-rotated with the hard-coded frequency difference. As a result of this, signals like DAQ_LNFS_56MHz_phi look flat. This phase is sent via TOLM at 10 kHz to the various photodiodes and SSFS_Ctrl processes, so that the signals can be demodulated at the correct angle even in case of small phase drifts.
The issue occurs when we change the modulation frequency: the phases of the LNFS are now no longer de-rotated with the correct frequency and will start wrapping around at +- pi, see fig 2. This is in principle not a problem, since adjusting a phase with a wrapped value still gives the correct answer. For the SSFS signal, however, the phase needs to be up-sampled to 500 kHz, so the anti-alias filter will cause a glitch every time the phase wraps around.
As a temporary solution, Alain removed the anti-alias filter of the phase in the SSFS_Ctrl process (so now the 10 kHz signals are simply repeated 50 times). As a more permanent solution, we could send both the LNSFS_*Hz_I/Q signals (which do not show any discontinuities, and can thus be safely up-sampled), and finally calculate the phase of the LO by doing a atan2 at 500 kHz.
We should also consider getting rid of the derotation step (which happens both at the level of the LNFS and the photodiodes), since the hard-coded difference frequency become arbitrary when we change the modulation frequency during every lock acquisition.
