nx_swarm_admit.nx

nx_swarm_admit.nx -- THE GENERAL ECOSYSTEM ADMISSION CAPABILITY (operator 2026-07-16: host- oversubscription control is the SWARM's job as the cluster scheduler, not a forge hack). Any workstream's heavy launcher (forge model decode, codec, image-gen, training, a big build) calls admit BEFORE starting; the swarm returns GRANT/QUEUE/DENY from LIVE host telemetry so the box never hits the load-105 collision (two uncoordinated model workloads). Supersedes the forge- specific nx_model_lane by generalizing it: workstream+kind keyed, telemetry-backed, one verdict. sa_verdict(load1, ncpu, avail_mb, est_mb, floor_mb, held, max_conc) -- PURE, gate-testable: DENY_LOAD if load1>=ncpu · DENY_RAM if avail<est+floor · QUEUE if held>=max_conc · else GRANT. Composes nx_model_lane_core (proven /proc telemetry: ml_load1/ml_ncores/ml_meminfo_avail_mb) -- DRY, no reinvention. RUNG 2 = fairness via nx_resource_arbiter (nx_ra_request) + auto cross-proc holder ledger; RUNG 3 = beacon nb_loadavg_milli canonical source + MCP-expose on the swarm plane. admit <workstream> <kind> <est_mb> <held> <max_conc> exit = verdict (0 GRANT/1 QUEUE/2/3 DENY) license_tier: ORIGINAL