sketch_kmv_jaccard_vs_naive_bench.nx

sketch_kmv_jaccard_vs_naive_bench.nx -- KMV Jaccard vs naive O(N). CLAIM TO VALIDATE: KMV sketch (Bar-Yossef-Jayram-Kumar-Sivakumar-Trevisan 2002, popular form Beyer et al. 2007) supports JACCARD similarity in O(k) memory regardless of set size. Naive baseline must materialize both sets, sort them, and compute |A∩B| / |A∪B| -- O(|A|+|B|) memory. Apache DataSketches ships Theta sketch with Jaccard, so KMV here is INTERNAL family completion + MEMORY axis vs naive. Both KMV and Theta are streaming approaches; naive is the materialized baseline. WORKLOAD: Set A: 1000 keys (ids 1..1000) Set B: 1000 keys (ids 500..1500) |A∩B| = 501, |A∪B| = 1500. True Jaccard = 501/1500 = 0.334. In PPM: 334000. MEASUREMENT: - KMV Jaccard estimate via nx_kmv_jaccard_ppm - Naive exact Jaccard from sorted sets - ACCURACY axis: closer to truth wins - MEMORY axis: KMV vs naive set storage HARD-WIN GATE: MEMORY BEATS by >50% KMV's Jaccard within 15% of truth