Fit the Conversation in the Window

Every chat product hits this on day two. The conversation grows, the window does not, and something has to go. Drop the wrong thing and the assistant forgets its instructions or the API rejects the request outright.

Write fit_history(messages, max_tokens).

messages is a list of {"role": ..., "text": ...} in chronological order, where role is "system", "user" or "assistant". A message costs len(text.split()) tokens. Return the messages to send, still in chronological order:

1. If the first message is a system message it is always kept, and it costs against the budget.
2. Fill the remaining budget with the most recent messages, walking backwards. Stop at the first one that does not fit: do not skip it and keep looking.
3. The kept window must begin with a user message. If trimming leaves an assistant message first, drop it too.

If even the system message exceeds the budget, return just the system message.