What Aggregate Scores Miss: Measuring Item-Level Regressions in Commercial LLM API Migrations