Transport-aligned robust reconciliation can degrade discrepancy localization: a controlled plant-network study

Conservation and model compatibility can produce coherent estimates without identifying which observation caused a discrepancy. A controlled sixteen-channel plant network compares weighted and Huber reconciliation, engineering thresholds, a trained causal histogram predictor and transport-aligned robust fusion with a fixed dry-solids relation. Twelve fault/control families, sixteen held-out seeds and four severity slots yield 6144 method/scenario cells. TASC recall is 0.3419063, compared with 0.3362188 for aligned Huber, 0.8399375 for the causal predictor and 0.5190938 after removing topology. The paired recall difference from no-topology is -0.1771875 with 95 percent seed-block interval [-0.1792188,-0.1752812]. Complete-frame methods also withhold more numerical entries, so coverage and source-valid comparisons accompany pooled ratios. Independent null-space optimization reproduces 48 robust objectives within numerical tolerance. The results distinguish optimization correctness, causal evidence timing and fault identifiability. They refute a general superiority claim for this particular equality/prior combination on the authored process, without showing that physical conservation is generally unhelpful or establishing operating-plant performance. Source code, executable lessons, data and computational records: https://github.com/fsantibanezleal/CAOS_OreBridge