Trimming patches with small dEs was a bust. They worked on Argyll but not I1Profiler which just says it can't make a profile. And even on Argyll the benefits were minimal and only occurred with about a 10% pruning.
So I examined the problem more closely. Turns out the 9800 has a very different problem than the 9500. The 9800 color variation is correlated with its location on paper and it is color dependent. By printing a large number of duplicate patches in each color then examining their statistics it turns out that colors exhibiting the largest variation are in the more saturated greens and, to a lesser degree, the oranges.
Specifically, out of a set of 70 distributed colors repeated 13 times in random locations, the mean dE00 of the worst case color, a fairly saturated green, was .38*. However the median color set of 13 patches was .17. Of the 70 sets of colors, 90% had a mean dE00 of .26 or less.
To mitigate this one can create a patch set where colors exhibiting the greatest variation are repeated N times as required to reduce the variance of the mean to the same levels as, say, just over the median color set. For instance 3 duplicates of colors closest to the worst case 10% color sets, and 2 duplicates for colors in the next 20% should significantly reduce the errors that will be baked into a profile.
These numbers may seem small but they are average dE00s. The variation across the 13 colors typically maxes around 3 times larger. So baking in fewer errors in the profile, which then compound with the unavoidable intrinsic error when printing seems desirable.
Another approach could be just printing two sets of charts with the colors scrambled differently. Then, after drying and scanning them, identifying the deviates and printing a third set of just those, perhaps duplicated once again and averaging the whole lot. Simple, but a lot of paper.