Hold violations can be worked around, so "different hold violations across runs" would mean that both sides pass but the distribution of independent and dependent hold violations changes. I have not seen that. The failures are something different---glitches in the run that are not being properly handled by the software. Part of the problem is that we don't have an understanding of these instantaneous failures.

Yes, Look like Software does has the ability skip the failing GPIO bit and re-adjust the shift chain depth and continue next search.