nx_forkgrade.nx

nx_forkgrade.nx -- GRADE THE WHOLE CROSS-TREE FORK, IN ONE PROCESS, IN NISHILANG. WHY THIS EXISTS. Every reading of the cross-tree fork so far was a SAMPLE -- 40 files, then 120, then 250 of 2,845 -- and the reason was purely instrumental: nx_srcdiff is one process per pair, and driven from the laptop each spawn crosses WSL at ~2s, so the population cost ~95 minutes and I shrank the evidence to fit the tool. ★★★★★★WHEN AN INSTRUMENT IS EXPENSIVE, PAY THE COST OR FIX THE INSTRUMENT -- NEVER SHRINK THE POPULATION TO FIT IT. Forking the same ruler NAS-side costs ~10ms, so the whole population is seconds. AND IT COMPOSES, IT DOES NOT REIMPLEMENT. The ruler stays nx_srcdiff -- one ruler, one implementation. This organ only iterates and tallies. A second line-differ would be the duplicate-ruler defect that this lane has already paid for twice (a local symbol-set check that was weaker than nx_srcdiff, and a second flow-back with a second guard). It also exists because the tally is JUDGEMENT, and judgement belongs in an organ: a .ps1 that classifies verdicts and counts distributions is exactly the shell-side judging the doctrine forbids. Shell's remaining job is moving bytes between the two trees, which is irreducibly laptop-side. usage: nx_forkgrade <listfile> <dirA> <dirB> [outfile] listfile : one tree-relative path per line dirA : NAS tree root (e.g. buildroot/runtime) dirB : staged other tree (e.g. /tmp/lapfork) outfile : per-file rows "<verdict> <bpaired> <bunpaired> <relpath>"; stdout carries only the envelope so a payload cap can never bound coverage. UNKNOWN IS ITS OWN BUCKET: a pair whose ruler output carries no anchored verdict is counted UNREADABLE, never folded into a known class. The bucket an unrecognised value lands in becomes the number somebody plans against. license_tier: ORIGINAL expect_exit: 0