sketch_reqsketch_vs_kll_bench.nx

sketch_reqsketch_vs_kll_bench.nx -- MATCHED-MEMORY paired measurement. HONESTY UPDATE (matched-memory rerun): Original bench used same k=64 for both; that gave ReqSketch ~3x memory (24 levels × 2*k = 3072 slots vs KLL 16 × k = 1024). This rerun MATCHES TOTAL SLOT COUNT: ReqSketch k=64 (3072 slots) vs KLL k=192 (16 × 192 = 3072 slots). Loser scaled UP, not winner nerfed down -- per substrate-honesty discipline. SHARED WORKLOAD: - Stream: integers 1..1000 in order - Slot budget: 3072 (matched) - Seed: deterministic, identical for both - Query: p99 -- the tail quantile where ReqSketch hra=1 should structurally beat KLL's uniform-eps bound EVEN AT MATCHED MEMORY (the structural-tail-preservation claim of Cormode et al. 2021). GROUND TRUTH: p99 of 1..1000 = 990 VERDICT REQUIREMENTS (honest, matched memory): ACCURACY p99 -> BEATS or EQUIVALENT (structural claim survives) ACCURACY p999 -> BEATS (gap should widen at deeper tail) ACCURACY p50 -> NOT BEATS (KLL's uniform bound holds at median) MEMORY -> EQUIVALENT by construction (matched slot count)