nx_bench_stress_autoreg_test.nx

nx_bench_stress_autoreg_test.nx -- 10-iteration REAL transformer stress test with wall-clock measurement. Where nx_actor_role_llm_v2_real_autoreg_test does 3 iterations for correctness, this smoke does 10 to exercise the substrate under load + capture honest wall-clock numbers. Each iteration drives the v2 LLM actor through INIT -> RUN (full Llama-class transformer forward via nx_llm_generate_one_v2) -> EMIT. Real embedding lookup + RMSNorm + multi-head attention + SwiGLU FFN + output projection + sample. Listener mailbox receives 10 LLM_TOKEN messages. Substrate invariants verified under load: * All 10 iterations succeed (every runner_result >= 0) * All 10 tokens are in [0, vocab) * Listener mailbox correctly handles 10 messages * Scheduler counters scale linearly (30 actor steps, 30000us cumulative runtime) * Wall-clock elapsed > 0 (substrate actually ran wall-time work) HONEST SCOPE: wall-clock under qemu-riscv64 isn't physical-hardware time. When Phase E + a physical-board arc land, this same smoke will produce defensible numbers without code changes.