KV cache

Explaining why long contexts cost more and why caching prefixes helps latency.

As a user-facing product feature name — say “prompt caching” instead.

After the system prompt is processed once, later tokens in the same session reuse cached KV blocks.