TBH I believe that your tests (which are really impressive, I adore the scientific approach!) confirm my assumptions…
Here you are showing that without optimized chia 1 plot takes maximum 3200 MB: Plotting process measurements - #4 by gladanimal
Here we see results for optimized chia which tops RAM at 4892 MB:
I understand that it is different hardware, but the impact shouldn’t be so huge. Also, in the first mentioned test you used 1 thread, and in the second two, and as we know even from your previous post, higher number of threads should rather decrease need for RAM.
And these tests are almost exactly in line with my “guestimate” observations that optimized chiapos uses in the worst moment 50% more RAM