Fix: handle non-string tool call arguments in memory consolidation

Fixes #1042. When the LLM returns tool call arguments as a dict or
JSON string instead of parsed values, memory consolidation would fail
with "TypeError: data must be str, not dict".

Changes:
- Add type guard in MemoryStore.consolidate() to parse string arguments
  and reject unexpected types gracefully
- Add regression tests covering dict args, string args, and edge cases

nanobot/agent/memory.py

@@ -125,6 +125,13 @@ class MemoryStore:
                 return False
 
             args = response.tool_calls[0].arguments
+            # Some providers return arguments as a JSON string instead of dict
+            if isinstance(args, str):
+                args = json.loads(args)
+            if not isinstance(args, dict):
+                logger.warning("Memory consolidation: unexpected arguments type %s", type(args).__name__)
+                return False
+
             if entry := args.get("history_entry"):
                 if not isinstance(entry, str):
                     entry = json.dumps(entry, ensure_ascii=False)