sketch_thompson_vs_ucb1_bench.nx

sketch_thompson_vs_ucb1_bench.nx -- paired measurement, Thompson vs UCB1. CLAIM TO VALIDATE: Thompson Sampling (sketch_thompson) achieves LOWER regret than UCB1 (sketch_ucb1) on a 2-armed Bernoulli bandit with arms at 80% and 20% over 1000 steps. Modern Bayesian-bandit literature (Chapelle-Li 2011, Russo-Van Roy 2018) reports Thompson winning by ~10-15% empirical regret on this setting. We measure the same on our substrate. SHARED WORKLOAD: - Same simulator LCG state seed (999) for both - 1000 steps each - Arm 0 = 80% (true mean 0.8), Arm 1 = 20% (true mean 0.2) - Optimal cumulative reward = 800 (always pick arm 0) MEASUREMENT: ACCURACY axis: cumulative successes (closer to 800 wins). MEMORY axis: nx_thompson_memory_bytes vs nx_ucb_memory_bytes. TIME axis: INCONCLUSIVE. HARD-WIN GATE: ACCURACY axis must BEAT by at least 1% (10000 ppm).