Grug says prompt caching just store KV-cache which is sequenced by token. Easy cut it back to just before edit. Then regenerate after is just like prefill but tiny.