nx_swarm_coord.nx

nx_swarm_coord.nx -- SWARM FABRIC class-aware COORDINATOR (SF-R-COORD): the missing UNIFIER that routes any job to the right COMPUTE CLASS across the heterogeneous pool. Fixes the "ignored the GPU" gap in the fabric AND coordinates the fabric with the EXISTING mesh -- it does NOT duplicate them: * GPU classes (image/video/render/train) -> the GPU worker roster from nx_worker_dispatch (wd_route/ wd_worker_where: gpu-image=laptop 5080, gpu-video=west 3090). Roles/endpoints are that organ's SSOT; inlined here (like nx_dispatch_lease inlined rt_lock) to keep ONE nx_syscalls surface + dodge the double-main trap of importing an executable organ. * CPU classes (cpu/tile/transcode/hash) -> the fabric's nx_swarm_place over live beacon telemetry. * CLOUD/LLM classes -> the cloud/no-float lane. * ALLOCATION under contention (VRAM/GPU-us/CPU-us fairness+preempt) = nx_resource_arbiter (already models these as first-class; its budgets are "fed by nothing" today -> the beacon telemetry is the FEED to wire). ★★THE LIAR-KILLER = THE OPERATOR'S CORRECTION, MECHANIZED: a GPU-class job is NEVER placed on a GPU-less node. If no GPU is available it QUEUES (or falls back to ANOTHER GPU) -- it must NOT silently run gen on CPU. The gate proves this can't happen. That is "don't ignore the GPU", turned into a test. route <class> <gpu_avail_csv> -- decide subsystem+target for a job class given which GPU roles are up [gate] -- self-gate: routing discriminates + the never-ignore-GPU liar-killers gpu_avail_csv = comma list of live GPU roles, e.g. "gpu-image,gpu-video" or "" (none up). In production this comes from the mesh probe / the beat store once GPU nodes beacon. license_tier: ORIGINAL expect_exit:0