Are AI-generated test cases reliable? Scenario fit ladder, false-coverage risks, human review gates, and historical bug-clustering as a generation strategy.