fix(llm): catch JSONDecodeError in LLM retry loop

Handle malformed JSON responses from LLM providers (e.g., OpenRouter
returning truncated JSON) by treating json.JSONDecodeError as a
retryable error. This allows 3 retry attempts with exponential backoff
instead of immediately crashing the run.

Previously, JSONDecodeError from the OpenAI SDK's response parsing
would escape all error handlers and kill the game run. Now it's
treated like APIConnectionError — logged and retried automatically.

Author: S1M0N38

src/balatrollm/llm.py

@@ -1,6 +1,7 @@
 """Async OpenAI client wrapper with retry logic."""
 
 import asyncio
+import json
 import logging
 from dataclasses import dataclass, field
 from typing import Any
@@ -113,6 +114,10 @@ async def call(
                 logger.error(f"LLM content filter error: {e}")
                 last_exception = e
 
+            except json.JSONDecodeError as e:
+                logger.error(f"LLM response parse error (malformed JSON): {e}")
+                last_exception = e
+
             if attempt < self.max_retries - 1:
                 logger.warning(
                     f"Retrying in {retry_delay}s [{attempt + 1}/{self.max_retries}]"