Use parallel subagents to scale large reviews, with an honest model of where the speedup actually comes from (LLM reasoning, not API throughput)
