Designing a Multi-Agent Review Pipeline for a Large PR

A 2,000-line PR touches authentication logic, a database migration, and a hot-path performance-sensitive function. You want to use a multi-agent review pipeline instead of one agent reading the whole diff top to bottom.

1. Design the pipeline: what subagents would you fan out for Stage 1, and why is dimension-based fan-out (rather than, say, splitting the diff into three arbitrary chunks) the right split for this PR?
2. Explain concretely what the adversarial verification stage is checking for, using one plausible false-positive scenario from this PR as the example.
3. Why does a single agent reading the whole diff serially tend to underperform this pipeline on a PR shaped like this one — be specific about the mechanism, not just "more agents is more thorough."