nx_leakscan_gate.nx

nx_leakscan_gate.nx -- STANDING DETECTOR for the allocate-without-free class. WHY THIS EXISTS (2026-08-01): two INDEPENDENT runaways collapsed the host in one night -- nx_docportal_admin_daemon (16.5 GB resident in 10 SECONDS, 842 GB of address space over 49h) and nx_ts_lumadiff (28.4 GB, OOM-killed 7 times). Both were found BY HAND, mid-outage, after the control plane had already died and three other instruments reported healthy. Nothing was watching the one signal that named them both instantly. THE SIGNAL: an organ that only ever grows has VmPeak == VmSize. A healthy process that allocates and frees has VmPeak > VmSize, because its peak was higher than where it sits now. That single comparison separated the four real leakers from jellyfin (VmPeak > VmSize, normal churn) on the first pass, with no tuning. ★ IT REPORTS ADDRESS SPACE **AND** RESIDENT, AND RANKS ON RESIDENT. VmSize alone cries wolf: a 842 GB VmSize with 602 MB RSS hurts Committed_AS but not the page cache, while 28 GB RESIDENT is what actually drives swap to zero and puts 78 processes into D state. A count without its classification cries wolf; report both, rank by the one that hurts. EXIT CONTRACT: 0 = no process over the resident budget. 1 = at least one OVER (a fix list, loudly named). 2 = the scan could not measure (refuses to report GREEN when it did not look -- absence of findings and absence of measurement must never render identically). nx_leakscan_gate [rss_budget_mb] default budget 4096 MB expect_exit: 0 license_tier: ORIGINAL