Why rescue-based transcript ranking can mislead: normalization displacement and shared-control coupling in perturbation transcriptomics

When a disease-associated transcript moves toward control after a phenotypically effective rescue, the pattern can be read as mediator evidence. We show why this inference can fail even when the computation is reproducible. Across normalization schemes in PLP1-mutant Jimpy brain (GSE277705), the original negative-interaction criterion classified 47.6-76.0% of all expressed genes and 77.6-86.6% of disease-eligible genes. Under total CPM, the excess was driven largely by a +1.016-SD median WT treatment response; median-of-ratios (MOR) reduced it to +0.154 SD while leaving the Jimpy median near zero. Under MOR, mirror enrichment among disease-increased (77.6%) and disease-decreased (74.0%) genes showed selection dependence but could not distinguish mechanical coupling from genuine state restoration. Slope decomposition gave a 48.7% mechanical-to-observed ratio (-0.125 of -0.256); deleting one of four untreated-Jimpy samples moved it from 37.8% to 90.1%. Candidate-local point estimates were 60.8% near TRIB3 and 38.6% near DDIT4; the TRIB3-minus-DDIT4 difference stayed positive in all four deletions (8.6--27.7 points), but its magnitude remained unstable. Oligodendrocyte-specific Perk deletion robustly suppressed TRIB3 (R^{JP}=-0.817, 95% CI [-1.048,-0.587]) and DDIT4 (-0.292, [-0.419,-0.169]), whereas 2BAct resolved neither transcript-specific rescue. Rescue transcriptomics can therefore establish broad state restoration without identifying which downstream transcript mediates it.