nx_perf_sentinel.nx

nx_perf_sentinel.nx -- the team's BOTTLENECK/LIVENESS organ. Operator 2026-06-10: "slowness anywhere in our entire ecosystem is a total failure, build the team to identify these bottlenecks and dead ends and infinite loops etc and handle them." What it does (one re-runnable pass; the loop runs it on a beat): 1. samples /proc twice (SENTINEL_WINDOW_MS apart) and computes REAL per-process CPU%% from utime+stime deltas -- no vibes, kernel truth 2. classifies each process against a BUDGET TABLE (config file, NO magic numbers in code -- Cardinal 11): SPIN cpu%% >= spin_pct (the nx_signaling 100%%-spin class, bd.t/ovw multi-day runaways) RUNAWAY cpu%% >= hot_pct AND age >= age_s (hot for days) 3. HANDLES per the budget action: log -> sentinel log line; escalate -> ESCALATE line (the PM/loop dispatcher's queue); kill -> nx_kill(pid,9) ONLY when the budget line explicitly says so (governed: default budgets never kill -- a human/PM line opts in) 4. exit code = number of violations (0 = ecosystem clean) so it IS its own gate and cron-able health check. Budget config (default /tmp/nishi_perf_budgets.conf, overridable argv[1]): # comm-substring spin_pct hot_pct age_s action(log|escalate|kill) * 97 80 86400 escalate nx_signaling 90 50 3600 escalate '*' = default row. First matching row wins; comm match = substring. Composes: nx_syscalls (getdents64/read/kill), the nx_loop_monitor verdict->handle pattern, nx_sysload's /proc-truth doctrine. license_tier: ORIGINAL