60 Subagents, 650 Ideas, One Proof: What Anthropic's Riemann Run Teaches AI Engineers