sketch_tdigest_v2_vs_v1_bench.nx

sketch_tdigest_v2_vs_v1_bench.nx -- paired measurement, T-Digest v2 vs v1. CLAIM TO VALIDATE: v2 (Dunning arcsin k1) uses fewer centroids than v1 (simplified parabolic 4q(1-q)) at comparable global accuracy. The memory efficiency is the headline T-Digest property; v2 is supposed to honor it; v1 over-allocates centroids near tails. SHARED WORKLOAD: - Stream: integers 1..2000 in order - Both alloc with delta=100 MEASUREMENT AXES: ACCURACY: median + p10 + p90 errors; both should hit ~1% rank; v2 may be slightly worse (looser tail) but must stay within tolerance band MEMORY: n_centroids_used. v2 < v1 expected. TIME: INCONCLUSIVE (no in-program clock) HARD-WIN GATE: MEMORY axis must BEAT v1 by >5% (50000 ppm). Accuracy on median must be EQUIVALENT (not LOSES).