Model choice is not just accuracy. For long code reviews, latency and context reset cost can decide whether people keep using the tool.