nx_sweepcheck.nx

nx_sweepcheck.nx -- MECHANISES THE MOST EXPENSIVE RECURRING ERROR OF THIS PROGRAM: reporting a result from a parameter that never reached the code. WHY THIS EXISTS (measured, not theorised). Across one session the same defect class was hit repeatedly: * a reference-weight sweep returned 523/517/518/523 -- flat, because the handler never parsed the argument. I nearly concluded the model change had failed. * an index-offset A/B read 504 vs 508 and was declared "noise". The knob was hardcoded downstream; the 4-point difference was PROMPT TEXT, not the knob. * a dilation sweep at reach 0 / 4 / 12 returned BYTE-IDENTICAL results and was written up as a regression WITH AN INVENTED ROOT CAUSE. The verb had fallen through to a different code path. In every case the tell was the same and free: A SWEPT PARAMETER THAT PRODUCES NO VARIATION IS NOT AN INEFFECTIVE PARAMETER -- IT IS AN UNWIRED ONE. Analysing such results is worse than useless, because a plausible story gets attached to numbers that mean nothing. Per the banked law "A SESSION MUST REMEMBER < A TOOL THAT CANNOT DO THE WRONG THING", the check is mechanised here instead of trusted to discipline. Feed it the measurements from any sweep BEFORE interpreting them; it refuses the ones that carry no information. nx_sweepcheck values <v1> <v2> ... -> GREEN if the sweep varies, RED if it is flat (unwired) nx_sweepcheck selftest expect_exit: 0 license_tier: ORIGINAL