fix(meta/stats): recompute quantiles from real data — per-episode aggregation made q01/q10/q50/q90/q99 far too tight (~30-60% of true spread); mean/std/min/max were correct. State/action quantiles now match the openpi pi05_droid checkpoint norm_stats.

mpark