sketch_ams_vs_naive_bench.nx

sketch_ams_vs_naive_bench.nx -- AMS F_2 vs naive exact tracker. CLAIM TO VALIDATE: AMS sketch (Alon-Matias-Szegedy 1996) estimates F_2 = sum_i f_i^2 in O(d*s) memory regardless of distinct-item count. Naive exact tracker needs O(distinct_items) memory. Apache DataSketches does NOT ship AMS or any F_2 estimator -- this is greenfield in the summarization-sketch family. WORKLOAD: 1000 distinct items with Zipfian-ish counts. MEASUREMENT AXES: ACCURACY: AMS estimate vs true F_2 (computed naively) AMS bound: rel stddev ~ sqrt(2/s) at conf 0.68. For s=256, rel stddev ~ 0.088, so |AMS - truth|/truth should be <= 2*sigma = 0.176 (95% conf). We allow 25%. MEMORY: AMS bytes vs naive bytes; expect AMS smaller for >100 distinct items. HARD-WIN GATE: MEMORY BEATS by >50% AND AMS estimate within ENV envelope band of truth.