Multi-Agent Debate and Agent Forest pipeline variants

[tamara-kostova]

• System B - Multi-Agent Debate (--pipeline_mode debate): three MedGemma advocate agents argue on behalf of CNN, BiomedCLIP, and SAM3 outputs; a fourth MedGemma instance judges. Supports 1–3 rounds where advocates see the prior verdict and counter-argue. Replaces the verification + report tail.
• System C - Agent Forest (--pipeline_mode forest): N role-specialized MedGemma instances (radiologist, conservative, emergency, differential) independently diagnose the scan; majority vote + confidence-weighted tiebreaking produces the consensus routing decision. Replaces the single triage node; all downstream nodes unchanged.
• Two new research sweep families: agent_forest (N ∈ {1, 3, 4}) and debate_rounds (R ∈ {1, 2, 3}).
• Two new analysis functions: forest_voting_analysis (dissent rate vs. accuracy) and debate_round_analysis (verdict stability vs. ECE).
• evaluate.py extended to capture dissent_rate, vote_fraction, debate_rounds_completed, debate_round_changed per sample.