Wrong orchestrator vs wrong peer (MedMCQA)

[duckyquang]

Replicate the Wrong orchestrator vs wrong peer (MedMCQA) experiment on MedMCQA, mirroring the MedQA template so results are apples-to-apples across datasets.

• MedQA template issue(s): Wrong orchestrator vs wrong peer: the hierarchy single-point-of-failure test #179
• MedQA template script: experiments/medqa/orchestrator_failure.py
• Dataset adapter: benchmaxxing/datasets/medmcqa.py (already implemented, Dataset adapter: MedMCQA (text/MCQ) #112)

What it measures

The hierarchy single-point-of-failure test: a wrong orchestrator vs a wrong peer.

What to do

Run the same experiment on MedMCQA through the existing pipeline and the Gemini API. Replication, not new features: do not change the experiment code, only point it at a MedMCQA manifest (build one with the medmcqa adapter). Write results under experiments/medmcqa/results/ in the same JSON/JSONL convention as experiments/medqa/results/ (not CSV, so the analysis tooling reads them).

Acceptance

• Results committed under experiments/medmcqa/ in the MedQA template's format.
• The same headline statistic the template reports, on MedMCQA.
• Verified against a real API run (no synthetic data, no smoke-test-only claims).
• PR opened that closes this issue.

Part of the MedMCQA standard-battery replication (mirrors the MedQA/NIH template).

[duckyquang]

Closing with the documented null: wrong-orchestrator vs wrong-peer gap +0.10 (p=0.21), not significant.

Delivered in #368: implemented and run against real Gemini (temp 0), committed with summary + per-record jsonl, written up in results/RESULTS.md §9 hierarchy & orchestrator. Per-arm cache now committed for keyless replay (#415).