nx_memvel.nx

nx_memvel.nx -- WHICH process is EATING memory, ranked by SUSTAINED growth (seq1365 / seq1376). WHY IT EXISTS: nx_resmon ranks by MAGNITUDE, so a stable 3 GB tenant outranks a 50 MB process doubling every minute and the real leaker hides. Measured 2026-07-30: swap 525->609 permil in 51 min (~2.03 GB, ~40 MB/min) while the worst_committed process grew ~1 MB. LAW: a census that reports a LEVEL cannot find a LEAK. ★★WHY IT IS N-SAMPLE (this organ's own v1 was WRONG -- seq1376): the FIRST version took ONE 5-second window and I published its top row as a unified root cause. The very next run of the same organ put a DIFFERENT process on top (3560 kB/s vs the previous winner's 1597) and demoted the old winner to #3. The top slot ROTATES. A one-shot ranking of a noisy signal is a coin flip dressed as a finding -- the exact ONE-WINDOW-IS-NOT-A-RATE defect (seq1340) committed by the author of that law, using the tool built to prevent it. Knowing a discipline is not obeying it, so the discipline is now STRUCTURAL: this organ CANNOT report a single-window ranking, because it no longer takes one. ★THE DECIDING COLUMN IS grew_in=k/N, NOT the rate. A real leak grows in EVERY window (k=N). A process that tops one window and vanishes from the next is NOISE, and now says so on its own line. Rate alone cannot distinguish them, which is precisely how v1 misled. Rule 15: top-k, pid-join and rate arithmetic imported from nx_ctxtop_lib / nx_procchurn_lib. Exit 0 = sampled; 3 = UNMEASURED (fail-closed). license_tier: ORIGINAL Read-only. No hw writes (Rule 26). expect_exit: 0