nx_swarm_arbitrate.nx

nx_swarm_arbitrate.nx -- SWARM FABRIC cross-workstream RESOURCE ARBITRATION (SF-ARB): the layer that keeps COMPETING workstreams (a gen job, LLM inference, model training) sharing the ONE heterogeneous pool FAIRLY -- no workstream monopolizes, none starves, high-priority preempts low. This is the orchestration north-star's R-ORCH-4 (competing-workstream arbitration) for RESOURCES, and it FEEDS nx_resource_arbiter (which already models VRAM/RAM/CPU/GPU allocation with max-share fairness + priority + preemption but was "fed by nothing"). COMPOSES (check-before-build 2026-07-15): nx_resource_arbiter (nx_ra_new / set_budget / set_max_share / request / preempt / total_held) -- verified compiles+gates today. This organ is the FEED (pool budgets from fleet totals = the beacon telemetry) + the workstream-request driver + the fairness gate. It does NOT reimplement allocation -- that's the arbiter's job (rule 15 DRY / rule 9 single-responsibility: pack places chunks on participants; THIS shares the pool across workstreams; the arbiter allocates). ★★THE LIAR-KILLERS: (1) NO-MONOPOLY -- a workstream requesting > max-share of a resource is DENIED_MAX_ SHARE (can't take the whole pool). (2) NO-STARVATION -- after the greedy high-demand workstream is capped, a lower one still gets a (partial) grant, never zero-by-design. (3) BUDGET-HONEST -- total granted never exceeds the pool. (4) PRIORITY-PREEMPTION -- a higher-priority workstream reclaims from a lower one. (5) determinism. nx_swarm_arbitrate -- self-gate: the competing-workstream fairness scenario license_tier: ORIGINAL expect_exit:0