Median and unscaled MAD of this snapshot's weighted centroid distribution. Every deviation uses the same median. Centroid compression loses within-centroid spread, so both estimates approximate the original observations. At an exact half-weight boundary the two adjacent values are averaged.

Uses O(k) time and O(1) working memory for k sorted centroids. Precomputed quantiles are not required. The result retains this snapshot as merge evidence.

Throws