Re-ran the full 12-library dreg-vs-pydreg benchmark on an NVIDIA Titan Xp now that the pmv_laplace tail-truncation fast mode is the default (0.3.0). Median walltime speedup 4.51x (was 3.76x on P100), median peak-RSS reduction 5.38x, median called-peak Jaccard agreement 0.999362 -- all consistent with or better than the prior P100 benchmark.
Add completed dREG G6 run under G6-tmp (dreg/G6.time.log recorded an interrupted run; this is a real rerun pending validation before promoting/replacing G6)
