Short write-ups, kept in the order they were useful rather than the order they were written.
Take the same capture, build one interval histogram from presentation timestamps and another from the arrival time of each buffer, and the two will not match. The first describes what the source intended; the second describes what the transport and the operating system scheduler did to it afterwards.
Neither is wrong. The mistake is averaging them together, or quoting one while having measured the other. Whenever a jitter figure is quoted without saying which clock it came from, it means very little.
It is tempting to normalise a sequence to a convenient rate first and analyse afterwards, because every tool downstream is happier that way. But interpolation invents samples that sit exactly where the interpolator expects them to sit, which flatters any statistic sensitive to short-term variation.
The habit worth keeping: measure on the native grid, then resample for display only, and label the plot with which of the two it is.
A disciplined reference is the correct answer and also an expensive one. For bench work over minutes rather than days, a free-running oscillator logged against a second independent source is usually sufficient, because drift over the measurement window can be fitted and removed afterwards.
The failure mode to watch for is a reference that shares a power supply, or a common upstream clock, with the thing being measured. Then the two drift together, the residual looks beautifully small, and the measurement is meaningless.
Half of these entries exist because a result was rediscovered for the second or third time. Writing it down is faster than deriving it again, and considerably faster than trusting a memory of having derived it.