Where Claude Code's tokens actually go, and how to keep long sessions affordable: prompt caching, context size, and the auto-compact window that decides when a conversation is summarised and why it sometimes keeps compacting.