sketch_hll_lgk12_vs_lgk8_bench.nx

sketch_hll_lgk12_vs_lgk8_bench.nx -- validates bits-up cap raise. SUBSTRATE FACT TO VALIDATE: NX_HLL_LGK_MAX raised from 10 to 12 this session. Adds alpha_m_sq_q64 entries for lg_k=11, 12 + Heule thresholds. Higher lg_k gives exponentially more registers (m = 2^lg_k) and tighter rel-stddev (1.04 / sqrt(m)). lg_k=8: m=256, rel stddev = 1.04/16 = 6.5% lg_k=12: m=4096, rel stddev = 1.04/64 = 1.6% (4x tighter) WORKLOAD: 100,000 distinct keys streamed into both sketches. Truth = 100,000. MEASUREMENT: ACCURACY axis: closer to truth wins. At matched-seed input, lg_k=12 should beat lg_k=8. HARD-WIN GATE: ACCURACY BEATS by >2x improvement.