sketch_ddsketch_vs_tdigest_tail_bench.nx

sketch_ddsketch_vs_tdigest_tail_bench.nx -- multiplicative-error tail. CLAIM TO VALIDATE: DDSketch (Masson-Rim-Lee 2019) provides MULTIPLICATIVE relative error on every quantile: |est/true - 1| <= 1/bpo. T-Digest provides empirical tail-tight ADDITIVE error. On a HEAVY-TAIL stream where the tail values span multiple orders of magnitude, DDSketch's multiplicative guarantee wins on extreme quantiles. WORKLOAD: 500 small values (in [1, 100]) + 50 huge values (geometric in [1000, 100000]) -- the heavy tail. Total N=550. True p99 ~ 60000 (top 1% are the huge values). True median ~ 50 (middle of small values). MEASUREMENT: DDSketch bpo=64 -> rel err 1/64 = 1.56% per quantile. T-Digest delta=100 -> ~1% global rank error, but at p99 of a heavy-tail the absolute value error scales with quantile value. HARD-WIN GATE: On p99 with truth ~ 60000: - DDSketch should be within ~2-3% (1000-1500 absolute) - T-Digest may be much further off due to centroid quantization at the tail ACCURACY axis BEATS by >5%.