nx_bench_scale_test.nx

nx_bench_scale_test.nx -- substrate scaling under increasing workload. Runs the v2 LLM actor at 3 different iteration counts (10 / 30 / 100) and emits ONE TSV row + wallclock_elapsed_us per scale to stdout. Each scale is an INDEPENDENT measurement: fresh session, fresh fixture, fresh PRNG seed, fresh wall-clock window. What this answers: * Does wall-clock scale ROUGHLY LINEARLY with iteration count? If yes -> substrate has no quadratic overhead under load. If no -> there's a scaling bug we need to find. * Do counters scale exactly linearly? (They should: 3 phases per iteration * N iterations = 3*N actor steps; 1000us logical per step = 3000*N us logical runtime.) Output format: standard nx_bench TSV header once, then one row per scale (labels "scale_v2_10" / "scale_v2_30" / "scale_v2_100") plus per-scale wallclock_elapsed_us lines. Operator can pipe to file + sort -t$'\t' -k7n to see the scaling curve. HONEST SCOPE: qemu-riscv64 wall-clock includes interpreter overhead; the SHAPE of the scaling (linear vs quadratic) is what matters, not the absolute numbers. Physical-hardware run will refine the absolute numbers; the shape claim should hold across both.