IntrinsicScan - Label-Blind Feature-Wise Intrinsic Binary Partition Scanning

Scans continuous molecular features for intrinsic binary partitions without using phenotype labels to determine the split. For each feature, observations are divided into two one-dimensional states by selecting the split that minimizes the total within-state L1 loss, defined here as the sum of absolute distances between each observation and the medoid of its assigned state. Thus, unlike a squared-error objective, deviations are measured according to their absolute magnitude rather than being squared. Binary phenotype labels are used only after the partition is fixed to orient the two states, quantify phenotype concordance using orientation-free Youden J, perform exact conditional inference, and optionally assess resampling stability or fixed-partition permutation calibration. The method is intended for preprocessed continuous omics measurements and does not imply that a fitted partition represents true biological bimodality or discrete biological states.