Mvp 1 agents

[Mildogrc]

This PR establishes the core multi-agent debate infrastructure (MVP 1), integrating state management, vector retrieval, parallel evaluation, and a command-line debate runner using LangGraph.

1. Core Orchestration & State Management (src/argument_lab/orchestrator/)

• Built out the core LangGraph structure (graph.py), defining the DebateState and the flow between the proponent, opponent, and evaluation nodes.
• Implemented robust LangGraph routing logic (looping, convergence/stalemate detection, and termination).

2. FAISS Ingestion & Retrieval (src/argument_lab/core/)

• Implemented faiss_index.py and a retriever.py wrapper to chunk and embed documents, allowing agents to ground their arguments in external evidence.
• Added a sample_corpus.json containing test documents.

3. Advanced Evaluation Engine (src/argument_lab/core/)

• Added evaluation.py and eval_prompts.py to handle 3 independent checks:
  • Judge: Scores arguments on consistency, support, relevance, and completeness. Detects convergence and stalemates.
  • Hallucination Checker: Verifies that cited evidence strictly supports the generated claims.
  • Contradiction Checker: Detects internal inconsistencies across an agent's prior arguments.
• Verdict Generator: A final node that synthesizes the entire debate history into a structured Verdict (winner, unresolved claims, summary) upon completion.
• Automated score penalties apply dynamically if hallucinations or contradictions are flagged.

4. Data Schemas (models.py, state.py)

• Introduced strongly-typed Pydantic models for argument components, strict flags (HallucinationFlag, ContradictionFlag), and the Verdict.
• Wired up LangGraph Annotated reducers in state.py to handle complex state accumulation.

5. Executable Setup Utilities (setup/)

• Added setup/ingest_corpus.py to allow users to build the vector index via CLI.
• Added setup/debate.py as the primary CLI runner. It streams state chunks round-by-round and prints beautiful terminal outputs.

6. Export & Persistence (exporter.py)

• Implemented export_debate to save the final graph execution to a structured JSON source of truth, along with a human-readable Markdown transcript of the debate.
• Updated README.md to reflect end-to-end setup and usage instructions.