Great stuff Nick. And thank you for sharing of course !
There's a word of caution though ...
With the clear notice that you announced part of it yourself already, I would be not so sure that you're looking at a signal decent enough to apply this analysis. This is two-folded :
1. The bandwidth of the analyser looks as if not high enough;
2. The grounding is the probe possibly is not in good fahsion.
This is how you won't receive a square, which is the signal coming from the clock.
This is how ground bounce may be injected in your analyser's signal.
In my view it can still be possible to observe as you do if only #1 is in order but #2 requires thorough real-time looking at the way the signal changes (when detected as error).
Of course, the "ground bounce" I'm talking about can be there internally (MoBo) just the same and this is how it requires real-time observation which I can't do (not being next to you I mean). So what it would require is "knowing" what can do it (from the MoBo's point of view) like for example a timer going on. Could be the Clock Resolution as we apply it in the software, but even with that in mind I don't think I would be able to do this because too indirect and complicated.
What I'm saying is that this wouldn't go for one such a "static" analysis, and that it merely requires a few different setups and then draw conclusions out of those. So like you suggested "one clock is better than the other". So, how ? And how does it show in the wave form ? And didn't you see already that by only changing the probes (detach and reconnect) changes the wave form ? I think you did.
Back to the bandwidth, I think you can also see how the shape of the individual cycles are (or can be) repetitive. Now this relates to the sampling rate of the analyser and here too, it requires real time observation. So think the analyser running on some clock rate and once in the xxx cycles all is the same again. Sadly the xxx can be millions, so again a tough thing to do (quite impossible, depending on what you see).
An example of what I mean could be the down-going slope in the first cycle in the rad circle, where the up-going peak will be detected as out of bounds. But now watch the remainder of the down-doing slope which is steeper than the others and now also runs out of bounds (near the bottom). And again in the middle of the dip where the next sample is now relatively higher because the previous one was captured too low. Just think that the average signal is represented, but the granularity of the X-axis (= time) is not high enough. How to combine this ? well, if the sampling rate (bandwidth) is not high enough, you'd get exactly this. First this peak is captured and because of that being at a too high level, the next one is additionally low (level). So because of the too low bandwidth the analyser detects an error here (could even be three for this one cycle) and you are happy with that ...
Of course you solve this by making the mask broader, as you btw already have done and now end up with 444 errors for the captured time period.
In the end you could be fooling yourself.
Additionally I'm pretty sure that if you set the mask broad enough, there will be no errors and then what. However, you could hope for a few per million which could be real error.
Lastly, I think I can tell you that no such thing as jitter can be observed from a wave form so deteriorated as this on one hand, and a time scale of 10ns (/div) on the other. Of course, you expressed it by means of error against the mask, but it is the same thing. However, still a bit dangerous to talk like this because such a "clock" can have 10s of ns of peak-peak jitter easily. But will depend on the specs for this (depicted by the MoBo) and you might know. Anyway the headroom you gave it looks to be ~5ns (black areas) and assumed the middle of that to be your target, any error implies 2.5ns deviation. And because it will happen at both sides of the set boundaries (other side for another cycle), it will be 5ns again, and more depending on the excursions.
Enough food for thought I think, and I can tell you that making a few comparisons like I suggested to make some better sense out of it, will require so much time (I think) that nobody is going to ask you to do it. But when you do I think we will be digesting it with much interest.
PS, hints :
a. Always pick the very closest ground point you can find, near the signal (solder a small wire if needed).
b. Make the ground of the probe wire as short as you can. This means : wrap it right around the prope itself until left just enough for making the ground connection.
c. If you can incur for less/more error just because of changing b. somewhat, draw the conclusion that you can never get there (because it can always be better though up till physical constraints).