The answer is not reused

Prompt caching does not paste an old answer into a new chat.

The new ending still changes what the model must respond to. The model continues processing from the cache boundary and generates a fresh answer token by token.

This is not a database of finished replies.

It is a head start on the input.