nx_axisvar.nx

nx_axisvar.nx -- DOES THIS JUDGE'S AXIS ACTUALLY VARY? A per-axis distribution census over a whole directory of renders, so a fitness built on those axes can be justified by measurement instead of by three hand-picked examples. nx_axisvar <dir> <prefix> <suffix> e.g. nx_axisvar /tmp bodyfit_ .png WHY IT EXISTS, MEASURED 2026-08-14: a body-evolution run improved 3 points in 10 generations and LOST to random search at equal budget. A 5-point sweep suggested the reason -- contour_axis sat at its ceiling and face_axis was binary -- but five renders is not a population, and changing a shipped fitness on five samples is the same error as tuning until a green appears. The search had already written ~132 renders; this reads ALL of them. ★★★IT REPORTS NUMBERS, NOT A VERDICT. An uncalibrated classifier must publish its distribution and let the reader judge. The ONE verdict it does emit is not a threshold but a FACT: an axis with distinct==1 took the same value on every single input, so it cannot discriminate anything, and including it in a mean dilutes every other axis while contributing nothing. ★NO SILENT CAPS. The directory walk LOOPS getdents64 until it returns 0 -- one call is a prefix of a directory, not a listing, and publishing that prefix's count as a total is how a census becomes a sample nobody knows is partial. The distinct-value table is bounded and says so when it overflows. ★A FAILED JUDGE IS COUNTED AND NAMED, never folded into the population: judged + failed must equal matched, and the sum is printed so the partition can be checked. exit 0 all matched files judged | 2 usage | 3 cannot open dir | 4 nothing matched | 5 some failed license_tier: ORIGINAL. No hw writes (Rule 26).