sketch_cpc_dense.nx

sketch_cpc_dense.nx -- CPC dense bitmap mode + HIP estimator. FIXES THE V1 SPARSE-MODE MEMORY LOSS. Sparse mode (sketch_cpc.nx) stores explicit coupons in a hash map: ~24 bytes per distinct coupon. At 1000 coupons: ~24KB, much larger than HLL_8's 128 bytes. That's why the v1 bench reported MEMORY: LOSES. DENSE MODE bitmap: m * w bits total (m columns x w rows per column). For matched accuracy with HLL_8 (m_hll=128, 9.2% stddev), CPC needs big_M ~= 128 coupons. Configurations: m=8, w=16 -> 128 bits = 16 bytes (8x smaller than HLL_8) m=16, w=8 -> 128 bits = 16 bytes m=32, w=32 -> 1024 bits = 128 bytes (matched accuracy as HLL_8x8) HIP ESTIMATOR (Cohen 2015, Lang 2017): On each NEW bit set (n_set transitions from k to k+1): kappa += big_M / (big_M - k) estimate = kappa COMPLEMENTS sketch_cpc (sparse mode): - sparse: small N, exact storage, larger memory per coupon - dense: bounded memory, HIP estimator, near-optimal variance LOSSLESS-LANGUAGE DISCIPLINE: rel_stddev ~ 1/sqrt(big_M), same as sparse-mode CPC. Production tier (bounded memory, deterministic).