nx_mesh_mux.nx

nx_mesh_mux.nx -- WORKER MESH: MULTI-MODEL multiplexing on one GPU (the Triton model-management analog). A single GPU worker can host MANY models (Z-Image, an LLM, SDXL, ...) but not all fit in VRAM at once. This is the RESIDENCY policy: route a request by model-id, and keep the hot models loaded within a VRAM budget, evicting the LEAST- RECENTLY-USED model when a new one must be loaded and there is no room -- exactly a VRAM-sized LRU cache of models (Triton's dynamic model load/unload + instance management). Composes with autoscale (load/unload) + the container registry (model artifacts) + spec-emit (a per-model worker). CLI: (no args) -> self-test GATE (routing + fit/evict/refuse + LRU correctness + never-over-budget) run <budget_mb> <id...> -> simulate a request sequence, print hits/loads/evicts/refused/peak-VRAM NO fake greens: the gate proves resident VRAM NEVER exceeds the budget, an oversized model is refused (not force- loaded), and eviction picks the LRU model (a later hit on a kept model proves it was not wrongly evicted). ORIGINAL