nx_worker_dispatch.nx

nx_worker_dispatch.nx -- WORKER MESH keystone: capability-scoped job dispatch across a sovereign GPU/compute mesh. The operator architecture: the NAS runs the orchestration container; it dispatches jobs BY API to compute WORKERS (this laptop's RTX 5080 for image gen, the west server's dual 3090s for video, the cloud for other work). Each job runs under a SCOPED CAPABILITY -- MCP-client style, it may ONLY touch the workers + storage it was granted, "trusted to do work" enforced deny-by-default. Results are pushed to consumable storage (nishifamily.com/gallery). This organ is the DISPATCH BRAIN + the capability enforcement. It: * ROUTES a job by type -> worker class (image->gpu-image / video->gpu-video / llm|embed|other->cloud) * ENFORCES the capability: a job is AUTHORIZED only if its caps grant BOTH the routed worker AND the storage target (missing either -> DENY). This is the "real MCP-type capability to access trusted resources". * RESOLVES the worker endpoint (ip:port:path) + the storage target (gallery ingest). * ACTUATES: for an authorized job it can CALL the live worker over sovereign TCP (nx_http_client) to prove the authorized call reaches real compute -- the first leg of NAS -> worker -> gallery. CLI: (no args) -> self-test GATE (routing + capability enforcement + neg-controls) probe <worker-class> -> live: GET the worker's /v1/models, report reachability dispatch <type> <storage> <caps> -> authorize (deny-by-default) + probe the routed worker + report Worker classes: gpu-image gpu-video cloud. caps = comma list e.g. "gpu-image,gallery". NO fake greens: the gate mechanically proves the capability check DISCRIMINATES (grant->allow, missing->deny). license_tier: ORIGINAL