feat: add council dissent analyzer with auto-deepening research

[dobosmarton]

Summary

When the Agent Council (multi-model scoring panel) votes on an idea, council members sometimes disagree sharply — one model says GO while another says NO_GO, or they diverge on individual scoring dimensions (e.g., Anthropic rates willingness_to_pay at 90 while OpenAI rates it at 50). Previously these splits were silently aggregated via median and majority vote, losing valuable signal.

The Council Dissent Analyzer detects these disagreements, triggers targeted follow-up research on the contested dimensions, then re-scores with augmented evidence — turning council splits from noise into a research signal for more informed decisions.

How it works

Initial council vote → detect dissent (spread > threshold per dimension)
  → if dissent found:
    → generate targeted follow-up queries via LLM
    → collect follow-up research (Tavily + Perplexity + Firecrawl)
    → re-score with augmented evidence via AgentCouncil
    → attach DissentAnalysis to final PreBuildScore

Changes

• New models (verdandi/models/scoring.py): DimensionDissent, DissentResolutionRound, DissentAnalysis — all frozen Pydantic models. Added optional dissent_analysis field to PreBuildScore.
• New config settings (verdandi/config.py): dissent_enabled, dissent_max_rounds, dissent_dimension_threshold, dissent_decision_split_required
• DissentAnalyzer (verdandi/agents/dissent.py): Core class with detect_dissent(), build_followup_queries(), run_followup_research(), resolve(), and _rescore_with_context(). Maps scoring dimension names to research dimension names for targeted queries.
• Scoring hook (verdandi/agents/scoring.py): After council.evaluate(), runs dissent resolution when dissent_enabled=True. Also adds mock dissent data to dry-run output.
• CLI + TUI display: Dissent analysis section in verdandi report and in the TUI experiment detail screen, showing contested dimensions, per-provider scores, resolution rounds, and whether the decision flipped.
• 19 new tests (tests/test_dissent.py): Detection, dimension mapping, model validation, resolution flow, and settings integration.

Verification

• ruff check: 0 errors
• ruff format: clean
• mypy --strict: 0 errors (94 files)
• pytest: 460 passed (19 new + 441 existing)