nx_pool_ring_gate.nx

nx_pool_ring_gate.nx -- regression witness for the RING ARENA fix in nx_thread_pool (2026-07-08). The task arena was a bump allocator (16384 slots); past that, nx_pool_submit returned -1 without queuing, and a caller that then delta-waited spun to the 2-billion timeout. A long-lived shared pool (the LLM forward submits thousands of tasks per token) hit this and wedged 428s/token. Fix: slot index wraps modulo arena_cap; SAFE because channel depth << arena_cap so a slot is consumed long before the ring wraps back to it. This gate submits N >> arena_cap tasks to ONE pool (forcing several full ring wraps), each atomically incrementing a shared counter, and verifies EVERY task ran exactly once -- the exact scenario that silently dropped tasks before. Checks: 1 shared counter == N (no task lost to a dropped submit) 2 pool completed-counter == N (delta-wait saw them all) 3 a SECOND batch on the same pool also completes (reuse across the wrap boundary) lineage_id: pool_ring_gate_v1