Build a miniature knowledge-update benchmark

The benchmark has four parts that stay separate: a generator that knows the answers, three functions that each turn (history, question) into a context, one answer prompt shared by all three designs, and a scorer that is exact wherever it can be. Every model call goes through one claude -p wrapper on your subscription. The decision that matters most is in the extract-then-store design: the update rule. A store that keeps every value with the date it was said, and shows the newest first, can answer both "what is the completion date now?" and "what were we working towards in October?". A store without dates can answer neither reliably.