Design a reviewed consolidation job

Consolidation is treated as a migration, not housekeeping: the model proposes a whole new store, code works out exactly what changed, and nothing goes live until two independent checks pass. Evaluation catches changes that make answers worse; the approval rule catches changes that evaluation can't see, such as deleted history that no evaluation question asks about. The run below shows both. The fake proposer makes one bad supersede (the client's confused "27 November" replaces the contract date), which evaluation catches, and four deletions of superseded facts, which the questions don't notice and only the rule stops.