nx_quad_real_emit_test.nx

nx_quad_real_emit_test.nx -- 4-modality REAL workload with wall-clock. Where nx_quad_companion_compose used the v1 synthetic 4-vocab LLM adapter, THIS smoke runs the production-shape 4-modality scene: NARRATOR v2 LLM (REAL Llama-class forward) priority 80 NPC v2 LLM (REAL Llama-class forward) priority 60 AUDIO_QC nx_audio_scene_grade (5-axis Q14 verdict) priority 50 GAME nx_tictactoe PERFECT vs PERFECT priority 70 All four share one nx_session. Wall-clock measured around the whole concurrent workload. TSV row + elapsed_us emitted to stdout. HONEST SCOPE: cooperative interleaving under qemu-riscv64, NOT physical-parallel kernel execution. But the substrate-honest claim PROVABLE today: one session hosts four heterogeneous modality actors each invoking a DIFFERENT other-agent primitive (transformer + audio + game) without state cross-contamination.