Function compact_with_llm 

pub async fn compact_with_llm(
    messages: &mut Vec<Message>,
    llm: &dyn Provider,
    model: &str,
) -> Option<usize>

Perform full LLM-based compaction of the conversation history.

Splits the message history into two parts: messages to summarize (older) and messages to keep (recent). Calls the LLM to generate a summary, then replaces the old messages with:

1. A compact boundary marker
2. A summary message (as a user message with is_compact_summary=true)
3. The kept recent messages

Returns the number of messages removed, or None if compaction failed.