Every request sends the same 20,000-token system prompt. What would prompt caching save us, and how could it make things worse?

Make your call first. Use the published answer and rubric to examine your reasoning when you are ready.