Message count-based compaction trigger (default 200) was a legacy safety
net from when token estimation was unreliable. With calibrated overhead
estimation using actual LLM PromptTokens, the token threshold alone is
sufficient — matching industry standard (Claude Code, Anthropic API,
LangChain all use token-only).
Also fixes EstimateTokens vs EstimateHistoryTokens inconsistency in
end-of-run overhead calculation (loop_history.go), and updates stale
CLAUDE.md docs from 75% to 85%.