nx_bench_scale_dist_emit_test.nx

nx_bench_scale_dist_emit_test.nx -- distribution measurement at 3 scales. Composes [[nx_bench_stats]] over multiple scale points to produce a real variance-vs-load curve. 30 total independent runs: * 10 runs of v2 LLM 10-iter -> distribution A * 10 runs of v2 LLM 30-iter -> distribution B * 10 runs of v2 LLM 100-iter -> distribution C Emits all three distributions side-by-side to stdout so the operator sees how min/max/p50/p95/stddev scale with workload. Substrate-honest claim this measurement supports: * If p50 scales linearly with iter count and stddev scales roughly linearly too -> substrate is well-behaved. * If stddev grows faster than mean -> there's a load-dependent variance bug. * If p95/p50 ratio stays roughly constant across scales -> no tail-degradation under load.