nx_mesh_autoscale.nx

nx_mesh_autoscale.nx -- WORKER MESH F3: sovereign AUTOSCALE-TO-ZERO controller (the Modal/KServe analog). A GPU worker should hold VRAM only while there is work. This is the LIFECYCLE POLICY: given whether the worker is warm (model in VRAM), whether there is demand (a pending/recent job), and how long it has been idle vs a TTL, it decides one action -- SPAWN (warm it), SERVE (dispatch now), HOLD (do nothing), or REAP (free VRAM). The policy is identical whether it runs natively on a NishiOS+VRAM host or drives a worker-host agent; the actuation (start/stop the worker process) lives at the host boundary (CreateProcessW on Windows today, native on NishiOS). Nishi is STRUCTURALLY ahead on scale-to-zero: idle = 0 MB / 0 procs (measured) and cold start ~1s, so demand-spawn + idle-reap costs almost nothing -- exactly where the serving frameworks are "Best". CLI: (no args) -> self-test GATE (decision table + savings + neg-controls) tick <demand> <idle_sec> <ttl_sec> -> live: probe the worker for warmth, then print the DECISION (a host agent calls this each cycle and actuates SPAWN/REAP) tick <demand> <idle> <ttl> <a> <b> <c> <d> -> probe an explicit worker ipv4 Actions: 0 HOLD 1 SPAWN 2 SERVE 3 REAP. NO fake greens: the gate proves every transition + that REAP never fires while warm-and-busy and SPAWN never fires while already warm. license_tier: ORIGINAL