State Before Action: Explicit Repair State Improves Multi-Operator Contradiction-Cluster Curation

Contradiction-heavy datasets fail at the cluster level: one keep/drop decision cannot simultaneously preserve legitimate ambiguity, recover hidden context, selectively relabel unsupported answers, and suppress annotation artifacts. We introduce StateRoute, a repair-state interface that lets the learned controller, budget allocator, and reviewer distinguish four causes of conflict before selecting among five cluster repairs under an explicit minority-collapse cost. A matched comparison fixes graph extraction, cluster evidence, action menu, controller-label pool, reviewer interface, and the $12.4$ min/1K review budget while contrasting explicit state with direct action imitation. Across five datasets, StateRoute reaches $0.82$ pooled pair-level contradiction F1, compared with $0.79$ for a GBDT action selector, 0.80 for a tuning-matched Cluster-Transformer selector, and $0.76$ for type-free graph repair; on MedQA, it reaches $61.7$ accuracy and $87.3$ minority retention, versus 60.4 and 61.0/82.5 for the two direct selectors. A disjoint 384-cluster action test localizes the gain to repair choice: StateRoute obtains 0.76 accuracy and 0.72 macro-F1, ahead of 0.70/0.65 for GBDT imitation and 0.73/0.69 for sequence imitation. The state advantage carries through all five Llama-3-8B downstream tasks, while Mistral-7B preserves the ordering over graph-only and judge baselines. Neutral-loss training, three retrieval/scoring backbones, leave-one-domain-out evaluation, cross-fitted controller supervision, and independent medical and legal-contract panels retain the advantage on their reported endpoints. Under this matched contradiction-cluster scaffold, operator-distinguishing state improves repair choice, downstream utility, and minority preservation beyond increasingly expressive direct action imitation, making the combined-system gain auditable under shared graph, evidence, and review-budget controls.