Fix prompt truncation with tool messages

[wmcgyver]

Summary

Fixes prompt truncation when tool messages follow the current user message.

When the prompt exceeds the context limit, the existing truncation logic can treat the current user message as old history and remove it. The following assistant tool-call and tool-result messages are then also removed, leaving messages empty and causing make_prompt() to fail at messages[-1].

This change tracks the most recent user message, only removes history preceding it, and truncates that user message directly when necessary.

Testing

• python -m py_compile modules/chat.py
• git diff --check
• Tested with a ~239k-token input truncated to fit a 32k context.
• Repeated successful web_search tool-call continuations after truncation.
• Tested failed fetch_webpage calls followed by successful model continuation.
• Tested with Qwen3.8-27B IQ3_M and IQ4_XS.
• No empty messages / IndexError occurred after the change.