nx_bench_v1_vs_v2_test.nx

nx_bench_v1_vs_v2_test.nx -- first paired substrate measurement. Per CARDINAL [[feedback-no-strawman-perf-comparisons]]: no perf ratio without paired measurement. This smoke is the smallest honest paired comparison: ARM A: nx_actor_role_llm (manual 4-vocab embed+matmul+sample pipeline, single token) ARM B: nx_actor_role_llm_v2 (real Llama-class transformer forward via nx_llm_generate_one_v2, single token) Both arms drive the SAME prompt + SAME PRNG seed through the same session shape (nx_session_new_default), then capture the substrate counters into NxBenchReport. The delta tells the operator: * How many actor steps each arm consumed * How much cumulative runtime each arm reported * How many messages each arm produced through the bus HONEST SCOPE: this is NOT a wall-clock benchmark. qemu-riscv64 is single-threaded; wall-clock comparisons are meaningless without physical-hardware runs (Phase E queued). What this primitive proves: the bench-capture substrate produces DETERMINISTIC paired counters, so when Phase E + a llama.cpp comparison rig land, the substrate already speaks the right language.