fix: guard against malformed LLM findings in conversation validation

Filter out non-dict entries from user_findings and handle non-dict
result to prevent 'str' object has no attribute 'setdefault' errors.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

utils/llm_client.py

@@ -514,7 +514,9 @@ class LLMClient:
     @staticmethod
     def _validate_conversation_result(result: dict) -> dict:
         """Validate and normalize conversation analysis result."""
-        findings = result.get("user_findings", [])
+        if not isinstance(result, dict):
+            return {"user_findings": [], "conversation_summary": ""}
+        findings = [f for f in result.get("user_findings", []) if isinstance(f, dict)]
         for finding in findings:
             finding.setdefault("username", "unknown")
             score = float(finding.get("toxicity_score", 0.0))