Code mode kept peak context to 749 tokens; the classic loop hit 78,628

Peak model context per run, estimated tokens, same 41-call task