fix: event classifier now strips markdown fences and repairs JSON

_parse_classification_response receives raw model output (with thinking
tags, markdown fences, etc.) but was calling json.loads directly.
Now uses _strip_markdown_fences + _repair_json from the client module
before parsing, matching what _call_ollama does for extractions.

services/extractor/event_classifier.py

@@ -284,9 +284,14 @@ def _parse_classification_response(
 ) -> GlobalEvent:
     """Parse raw Ollama JSON output into a GlobalEvent.
 
+    Strips markdown fences and repairs malformed JSON before parsing.
     Normalizes enum values and clamps numeric fields.
     """
-    data = json.loads(raw_json)
+    from services.extractor.client import _strip_markdown_fences, _repair_json
+
+    cleaned = _strip_markdown_fences(raw_json)
+    cleaned = _repair_json(cleaned)
+    data = json.loads(cleaned)
 
     confidence = data.get("confidence", 0.5)
     if isinstance(confidence, (int, float)):