Refactor narrative_nodes.json: 57.7% line reduction via templates, merged actions, inline end nodes

[Copilot]

narrative_nodes.json had grown to 1139 lines in prototype stage due to redundant per-action five-level result sets, 4 near-identical fire distraction actions, and 5 content-free _end placeholder nodes. This PR restructures the data and updates all consuming code.

Data file (narrative_nodes.json): 1139 → 482 lines (−57.7%)

New top-level structure:

{
  "outcome_templates": {
    "fire_distraction": { "大成功": {...}, "成功": {...}, ... }
  },
  "nodes": [ ... ]
}

• Merged fire actions: 火焰箭 / 火球术 / 点火术 → single 点燃类 action referencing shared template via {"$template": "fire_distraction"}. Key matches action name directly — no fallback lookup needed.
• Removed 5 _end nodes: altar_site_001_end etc. replaced by "__end__" sentinel in next_nodes + end_summary field on the parent node.
• Compact results_pool format: outcome level as dict key instead of list of objects, effects inline — eliminates ~5 lines per result down to 1.

narrative_system.py

• NarrativeNode.from_dict accepts templates: Optional[Dict] param; detects and handles three results_pool formats: legacy list, new compact dict, and {"$template": "..."} reference.
• New _resolve_template static method for {param} substitution in template text.
• load_from_file detects new {outcome_templates, nodes} top-level dict vs. old flat list (backward compatible).
• Added end_summary: Optional[str] field to NarrativeNode.
• NarrativeResult docstring documents the full effects field schema (30+ effect types) as canonical reference.

scene_manager.py / game_view.py

_handle_narrative_confirm now checks next_node_id == "__end__" to return to map, replacing the old is_end_node lookup against a now-removed node object.