I found a particularly egregious firmware error. That error pretty convincingly explains the 4 ppb difference observed against the calibrated signal generator at work. I thought 4 ppb was pretty good, but it may be even better when I try it again tomorrow.

I've analyzed the power bus some more. It appears that the majority of the ripple and noise on it may be coming from the oscillator. It's too late for the current run of boards, but I am going to try adding a ferrite bead between the two decoupling caps on the Vcc pin to see if I can squelch some of that. If I put a 50 MHz low-pass digital filter in place on my scope I see absolutely nothing at all on the 3.3 volt rail, which is kind of what you'd expect from using an LDO. We'll have to see how much garbage comes out of my buck converter design.

I expect the first batch of "production" boards towards the end of the week.