Upload & Sell: On
I think that decentering could possibly affect things, but in my experience decentered lenses affect autofocus accuracy rather randomly. I've seen cases where a lens simply misses more, but more specifically have seen zooms that are decentered in a focusing group so that they look good at one focusing distance but are decentered at another. These tended to miss focus randomly at the bad distance, but didn't exhibit the behavior described.
Here is another thought:
Phase is 'semi closed loop' and more accurate at saying '12 more feet over there' than 'here we are'. (Contrast detection is more accurate at saying 'here we are' and not at all accurate at saying even what direction to go in. Phase is more like a GPS, triangulating two different points and extrapolating where it should go. As it gets very close, the range of error is increasing. So it takes a couple of readings on the way to where it's going, but it's final 'here we are' is not accurate like contrast detection would be.
Each lens has a slightly different motor (we know they are calibrated for voltage and frequency in firmware), slightly different resistance to movement, etc. The resistance and movement are not linear, they vary according to a mathematical curve at different distances.
If the numbers for the lens are off slightly in firmware, it could affect it more at different distances.