nx_mesh_lb.nx

nx_mesh_lb.nx -- WORKER MESH: multi-worker LOAD BALANCER + HEALTH-AWARE routing (the Triton/Ray replica-routing analog). A worker CLASS (e.g. gpu-image) has N registered endpoints (laptop 5080, west 3090, ...). On a request the balancer HEALTH-PROBES each endpoint (bounded, non-blocking connect so a down replica is skipped in ~1s), then picks a HEALTHY one by round-robin (persisted counter -> spreads load), skipping every unhealthy replica. If none is healthy it refuses. This is what turns "one worker per class" into a real load-balanced mesh, and it is exactly how a second worker (west 3090) or a cloud replica joins: add an endpoint row, zero code. CLI: (no args) -> self-test GATE (round-robin + health-skip + no-healthy + liar-kills) pick <class> -> live: probe the class's endpoints, print health vector + the chosen replica Classes: gpu-image (laptop 5080 + west 3090). NO fake greens: the gate proves it NEVER picks an unhealthy replica. license_tier: ORIGINAL