refactor(llm): robust JSON parsing + centralized prompt engine

Three-layer defense for LLM JSON output reliability:

1. JSON parser rewrite (json_parser.py):
   - 4-stage repair pipeline: direct parse → control-char sanitize →
     json_repair library → structural repair for truncated output
   - Removed fragile _repair_unescaped_quotes heuristic in favor of
     json_repair library which handles ~95% of LLM JSON defects
   - Added json-repair>=0.30 to core dependencies

2. Centralized JSON instruction prompts (router.py):
   - _JSON_RULES constant with explicit quote-escaping guidance
   - _SYSTEM_TOOL_RESPONSE, _SYSTEM_JSON_ONLY, _SYSTEM_ROUTE
   - All 3 system prompt paths now include backslash-escape examples
   - Critical for small/medium models (7B-14B) that embed narrative
     dialogue with unescaped double quotes

3. Single parsing entry point:
   - All LLM→JSON parsing routes through _extract_json_object()
   - Consistent behavior across route(), generate_json(), and
     _generate_tool_response()

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: dennys246

pyproject.toml

@@ -30,6 +30,7 @@ dependencies = [
     "scipy>=1.15,<2.0",
     "llama-cpp-python>=0.3.8",
     "ddgs>=6.0.0",
+    "json-repair>=0.30",
     "pyyaml>=6.0",
     "openai>=1.0.0",
 ]