I looked at the CC PLL again. It still seems to be quite unstable. There was one occasion where I could get it to lock for more than 10 minutes. During phase noise measurement I saw that the noise was glitch type rather than stationary, where the noise floor could be at the level of previous measurements but quickly gets pushed up by an impulse excitation.

This time I measured the phase noise by providing 7 MHz as the local oscillator from DDS3 DAC0 (PPol LO @ 9 dBm), which corresponds to the usual operation frequency of the CC PLL beat note (7MHz at "-23.5" dBm on the "17dB more is measured " spectrum analyser,  DDS signal to CC LO is 21 MHz). Changing the LO offset to both 6.9990 MHz and 7.0001 MHz (+/-100 Hz offset) resulted in a corresponding 100 Hz signal output from the mixer. On the oscilloscope I can still quite frequently see the signal glitches.

I tried twice to get the phase noise spectrum but it seems the signal is not cooperating. The noise floor gets raised quite high by random noise impulses (figure 1). Qualitatively, the noise floor should be about at the level of the other measurements.

Figure 2 shows the Apk calibration to go from Vrms to rad/rtHz. Last time it was 0.0086 mV, now it is 0.0078.
Figure 3 and 4 show some feedback on the CC spectrum analyser at 14 and 28 MHz. The CC PLL is going to a T attached on the spectrum analyser. The other exit of the T then goes to the RF port of a mixer. When disconnect the mixer the 14/28... peaks go away.
Figure 5 shows the period of 2 wavelengths when the local oscillator frequency is offset to 7.000 100 MHz in DDS3 DAC3. So a 100 Hz difference between DDS and beatnote produces the 100 Hz signal as expected.