Caching expectations like that probably starved the optimizer of enough fresh data points to keep moving, especially once the landscape flattened. Batching a few circuit runs together before handing back might trim the handoff overhead without freezing progress.