Well, not to beat a dead horse, but just re-interpreting my original measurements, the result is exactly what one might hope for in the utility of the vision encoder plate:
Explanation: the original author’s method is similar to the most accurate method of measuring the distance between two pins. For those not familiar with that method, I’ll let chatgpt explain, and then I’ll continue afterward:
chatGPT:
Me again:
So, how does that relate to the original author’s model and method of measuring? You can think of it like measuring the distance between 2 pair of 2 pairs of pins, where the centers of the basic pin centers are spaced exactly 145mm apart (i.e. what it would be if you measured it in the original CAD model). The original author provides two such pairs for measurement in X, and the same for Y, presumably on the theory that averaging will help compensate for inevitable measurement error.
So, when looked at that way, only the original Y measurement (prior to the vision encoder plate calibration) was out by more than 0.03mm. After the vision encoder plate calibration, though, both X and Y are within that 0.03mm tolerance that one would expect.
Looked at this way, all is good! The vision encoder plate, at least in this instance, achieves what it claims. 
Granted, this is only one datapoint, but if the rest of you run the same or similar kind of measurement, and post your results, as I have, then by pooling our datapoints we can have more or less confidence, depending on how your measurements turn out on your respective machines. So, I would encourage anyone who is not a hopeless sponge to do so. The effort required is quite minimal. Just download the original authors model (I linked to it earlier in this thread), and post your measurements. The measuring process is easily less than 2 minutes tops.
Or, if you prefer to print a more thorough calibration model that casts a wider net, even better. More power to you.
P.S. I should perhaps add that the two models I printed were both the same third iteration print, where I had incrementally improved the shrinkage compensation factor for Sunlu HS-PLA that I used in the print to 99.7932% i.e. I used the same shrinkage compensation factor for both the “before” and “after” prints. As you would expect, not much shrinkage with PLA, which is partly why it is so popular. The net effect is that whatever I now print using that shrinkage compensation factor for that filament on the H2D (assuming the filament is equally dry) will be dimensionally accurate in X and Y. That’s also why I keep the humidity in my AMS2 very low: 0% as reported by the AMS2
though more likely 7%RH, as reported by the Yolink and Switchbot RH sensors that I sealed in the AMS2 as a check on the accuracy of the AMS2’s reported humidity level. Most likely Bambu used a cheaper, less accurate humidity sensor for the AMS2, maybe one that doesn’t measure at all below 10%,
but that’s a different topic for another day on some other thread.