fix: strip markdown fences in persona JSON parsing + update usage test

Personas often return JSON wrapped in ```json fences which caused
json.loads to fail and fall back to {"raw": text}. Now strips fences
before parsing, so persona outputs are structured dicts in TurnContext
and _summarize_slot can show meaningful previews in discussion log.

Update test_usage_integration: "hello" is now correctly classified as
simple → haiku (agent-manager enabled by default).

715 tests passing.

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

Author: Dave Leal

tests/test_usage_integration.py

@@ -68,4 +68,5 @@ def test_project_message_records_mention():
             event, say=MagicMock(), channel_name="alpha-dev"
         )
 
-    mock_record.assert_called_once_with("api", "claude-sonnet-4-6")
+    # "hello" is classified as simple → haiku by agent-manager
+    mock_record.assert_called_once_with("api", "claude-haiku-4-5-20251001")

tools/personas/__init__.py

@@ -9,6 +9,7 @@
 
 import json
 import logging
+import re
 from abc import abstractmethod
 from typing import ClassVar
 
@@ -58,8 +59,10 @@ def run(self, inputs: dict[str, dict]) -> dict:
         resolved_model = MODEL_MAP.get(self.model, self.model)
         msg = self._call_api(system, user, resolved_model, self.max_tokens)
         text = msg.content[0].text.strip()
+        # Strip markdown code fences (```json ... ```) before JSON parsing
+        cleaned = re.sub(r"^```[a-z]*\n?", "", text).rstrip("`").strip()
         try:
-            result = json.loads(text)
+            result = json.loads(cleaned)
         except (json.JSONDecodeError, ValueError):
             result = {"raw": text}
         result["_usage"] = {