nx_dual_real_llm_compose_test.nx

nx_dual_real_llm_compose_test.nx -- TWO v2 LLM actors concurrent on REAL transformer forward, with wall-clock measurement. Substrate-honest version of [[nx_dual_companion_compose_test]] which used synthetic 4-vocab manual matmul. THIS smoke runs two independent v2 LLM actors that each drive the FULL Llama-class runner (embedding + RoPE + RMSNorm + multi-head attention + SwiGLU FFN + output projection + sample) cooperatively interleaved through one nx_session. Each generates 5 tokens autoregressively. Output from both actors flows through one multiplexer to one listener. Substrate invariants verified under DOUBLED real-transformer load: * Both actors complete cleanly across 5 iterations each * 10 LLM_TOKEN messages delivered (5 from each) * All 10 tokens in [0, vocab) * Counters scale: 30 actor steps total (5 iters * 3 phases * 2 actors = 30) * Wall-clock measured around the whole concurrent workload * Multiplexer correctly orders chunks from both actors by priority HONEST SCOPE: cooperative interleaving under qemu-riscv64, not physical-parallel kernel execution. Phase E required for that. But the substrate-honest claim PROVABLE today: one session hosts two independent REAL transformer-forward actors without state cross-contamination, message-bus collision, or arbiter starvation.