Function build_compression_content 

pub fn build_compression_content(conv: &Conversation) -> (String, usize)

Build content for LLM compression.

Strategy: keep the last KEEP_MESSAGES messages at full fidelity, compress everything before that into one-line-per-round summaries. Returns (compressed_text, number_of_messages_to_remove).

This operates at MESSAGE level, not turn level, because turn_tracker counts user messages (1 user msg = 1 turn) but a single user message can produce 15+ LLM calls with 35+ messages.