nx_procchurn.nx

nx_procchurn.nx -- the PROCESS-CHURN axis the health plane never had (debt seq1317 / seq1318). WHY THIS EXISTS: 2026-07-30, nx_health said {overall:OK, degraded:0, down:0} with all 14 services UP while the NAS was forking 78.3 processes/sec sustained, burning 231 permil of an 8-core box in the KERNEL against 98 permil in userspace (a 2.36x kernel:user inversion) and taking 91718 context switches/sec. nx_resmon owns MEMORY, nx_netobs owns NETWORK -- the process-churn axis had NO organ, so the largest CPU consumer on the platform was invisible to every instrument we owned and was found by hand. This is that instrument. LAW: liveness is not performance. A board that only asks "is it up" cannot see the machine burning. WHAT IT MEASURES (two /proc/stat samples bracketing a real window -- deltas, never since-boot totals, because a since-boot average hides a storm that started an hour ago): forks/sec -- `processes` delta. THE headline: fork+exec is the most expensive thing a kernel does. ctxsw/sec -- `ctxt` delta. Rises with churn AND with thrash; corroborates, never diagnoses alone. intr/sec -- `intr` delta, first field. Device pressure, reported for context (not thresholded). kernel:user -- system:user+nice jiffies as permil. 1000 = parity. THE inversion detector. busy/idle -- share of measured capacity, self-normalising (no HZ and no core count needed -- every share is a fraction of the SAME jiffy total, so the arithmetic cannot drift when the box changes size. That is what makes this instrument hardware-portable). Policy lives in nx_procchurn_lib.nx so the gate exercises the SAME code (rule 9 + rule 15); thresholds are data-driven from knowledge/status/procchurn.conf (rule 11). Read-only. No hw writes (Rule 26). Exit code IS the verdict so a caller can gate on it: 0=GREEN 1=AMBER 2=RED 3=UNMEASURED. license_tier: ORIGINAL expect_exit: 0