At least for my scale problem, I did not move it terribly rapidly. My scale didn't miss counts in one direction, but it missed them in the opposite direction, resulting in zero drift. But it was confined to a region near the center of the table. (Did not observe the problem near the ends of the table.) There was crud on the head and the glass scale. Cleaning helped a little but did not cure the problem. I had a 0.030" zero shift every time I went +2 inches and back (after cleaning). Before that it was much much worse. A new head fixed the problem.
The seals were not that good and they allowed contaminants in.
Yes - everything you mention about getting things clean is important, and will affect the repeatability performance, but the symptom of 0.030" every time you reverse, regardless of where on the scale you happen to try it is a great diagnostic! It pretty much exonerates cleanliness as the core of the problem. It is also, in my thinking, more than an order of magnitude unacceptable!
The clarity of the logic signal(s), the logic threshold, and various kinds of signal noise mechanisms are at play here. Sure, get the spacing to the head optimum for the head focus. Pay attention to the head itself, and any plastic bits up to the photodiodes. IPA is your friend!
For the scales to work counting forwards and backwards flawlessly, there are two scales of ON/OFF contrasts, offset from each other by 90°phase, such that a HIGH transition from one is followed by a HIGH transition from the other halfway along the duration of the first. This allows the logic to sense a direction reversal. Some systems double the resolution by sensing the edge transitions. This works, but is more noise prone. A perfectly good scale logic signal can get messed up simply by plugging it into a display using a OV on a different ground path.
The whole subject of noise, interference, grounding, screening, avoiding common mode grounding paths, and the coupling of (high) voltages from every nearby mains lead into logic circuits is huge, and the number of ways the signal can get messed up expands exponentially, but there are definite good things that can be done.
One thing I try never to do is attempt to use any part of a braided screen as the OV return for instrumentation kit mounted on large mass metal. Connect a screen to the metal ground at one end only, and have all currents there and back stay within the screen. The 0V of a digital circuit should possibly, or maybe inevitably be connected to a earthed ground at only one place, if it has to be at all, and that place should be at the 0V of it's DC power source. However noisy is the kit around it (welders, motors, etc.) it is possible to bring back signals completely clean.
