Causal inference for group-contaminated structured outcomes: observable quotients, lossless reduction and exact randomization inference