fix: remove START/END prints from run_episode, convert to STEP format

inference.py

@@ -239,7 +239,7 @@ def run_episode(env: CodeGuardianEnv, client: OpenAI, MODEL_NAME: str, agent_sta
     episode_history = []
     step_count = 0
     
-    print(f"START: Episode {agent_state.episode + 1} - model={MODEL_NAME} task_id={obs.task_id} difficulty={obs.difficulty}")
+    print(f"STEP: Episode {agent_state.episode + 1} started - model={MODEL_NAME} task_id={obs.task_id} difficulty={obs.difficulty}")
     logger.info(f"[EPISODE START] Episode {agent_state.episode + 1} - model={MODEL_NAME} task_id={obs.task_id} difficulty={obs.difficulty}")
     
     while not env.done:
@@ -310,7 +310,7 @@ def run_episode(env: CodeGuardianEnv, client: OpenAI, MODEL_NAME: str, agent_sta
             break
     
     # Final episode log
-    print(f"END: Episode {agent_state.episode + 1} - final_score={episode_reward:.3f} total_steps={step_count} strategy={agent_state.prompt_strategy}")
+    print(f"STEP: Episode {agent_state.episode + 1} complete - final_score={episode_reward:.3f} total_steps={step_count} strategy={agent_state.prompt_strategy}")
     logger.info(f"[EPISODE END] Episode {agent_state.episode + 1} - final_score={episode_reward:.3f} total_steps={step_count} strategy={agent_state.prompt_strategy}")
     
     return episode_reward, episode_history