Panel review loop

The real-world corpus is intentionally untracked. Before its first use, follow the benchmark setup and run npm --prefix benchmarks run download-fixtures.

1. Define the user-visible surface and success criteria.
2. Select representative projects from benchmarks/fixtures/real-world/.
3. Capture actual output for each project.
4. Run panel-review on the evidence, not on an intended answer.
5. Implement the smallest coherent improvement that addresses consensus.
6. Re-run the same corpus and compare behavior.
7. Stop when the panel has no blocks and further changes do not materially improve the surface.

Keep the corpus stable across iterations. Preserve output contracts unless the plan explicitly approves a versioned change.