When you branch a conversation in CC, when pressing Esc twice for example, a new JSONL gets created that is very similar to the initial one.
However, obviously it doesn't have the same cost associated with it (previous output tokens are becoming input tokens for example). The combination of message id + request id of assistant message should only be counted once globally, as it heavily skews the numbers otherwise.
You can test that behavior by branching a very long, output intensive conversation multiple times and calculating the price, it's way more than it should be. Only the delta (new assistant message that contains the previous output as input) in the new conversations should be included.