Stop Caching the Whole LLM Response. Cache the Embedding.

Exact-match response caches hit 4% of the time. Embedding-keyed caches hit 60%. Here is the 70-line implementation and the cost-shape that justifies it.