I'm getting results like this at the moment:

That's not tremendously better than it was doing before. The firmware at the moment is adding a fractional value to the detected error (expected cycles - actual cycles over a 100 second period) based on the average reading of the phase detector over the sample period. When I tried making the phase error scale to ±1 error point (equivalent to 1 ppb), the result was less stable rather than more. But to be fair, I was running it outside of the chassis, which will lead to instability as well.

I think I'm going to try cutting the sample window in half and increasing the influence of the phase detector. Meanwhile, it's working well enough to share the design and firmware here and on github, so look for those updates shortly.