The card quoted a KS statistic against the IID 5% critical value. --speed-bins retargets the draw at the most deficient bin and discards surplus successes, so the bin counts have zero variance where an IID sample's would be multinomial, and the threshold has the wrong null. The counts are exact by construction and say nothing about how many attempts each bin cost; the card now says that instead.
fix(meta): task_index statistics named the source's index, not the merged one
Every episode's episodes_stats.jsonl entry claimed task_index 0 because the merge renumbered the parquet column without moving the statistics OF it. 14,200 of 14,400 episodes were affected; only the first family happened to be right. Recomputed from episodes.jsonl + tasks.jsonl and checked against all 14,400 parquet columns.
