Narrative arbitration: LLM-generated conflict resolutions

[theinterneti]

Part of #270

For each detected conflict: generate a narrative description that accounts for both influences and produces a coherent canonical resolution. Output: (resolved_state, narrative_description) pairs.

Acceptance Criteria

• NarrativeArbitrator.resolve(conflict_record, world_context) → (resolved_state, narrative_text)
• Prompt includes: both conflicting states, entity history, world rules
• LLM output parsed into structured resolved_state dict + narrative paragraph
• Batch processing for multiple conflicts in single LLM call where possible
• Unit tests with mocked LLM; narrative quality reviewed manually