fix: 3 fixes for Step 2, Step 4, Step 5

1. Step 5: Use empty model_config_dict={} so camel-ai doesn't spread
   api_key into create() - AsyncOpenAI reads env vars automatically.
   Step 3 unaffected (same env vars, just cleaner ModelFactory call).

2. Step 2: Fix Thai text truncation - isalnum() stripped Thai chars.
   Use re.sub(r'[^\w]', '', username, re.UNICODE) instead.

3. Step 4: Move get_language_instruction() to START of prompt (not end)
   and strengthen wording with MUST/IMPORTANT prefix.

backend/scripts/run_twitter_simulation.py

@@ -454,18 +454,13 @@ class TwitterSimulationRunner:
         
         print(f"LLM配置: model={llm_model}, base_url={llm_base_url[:40] if llm_base_url else '默认'}...")
         
-        # Pass api_key and base_url via model_config_dict
-        # camel-ai extracts these for the OpenAI client constructor
-        model_config = {}
-        if llm_api_key:
-            model_config["api_key"] = llm_api_key
-        if llm_base_url:
-            model_config["base_url"] = llm_base_url
+        # AsyncOpenAI reads OPENAI_API_KEY and OPENAI_BASE_URL from env automatically
+        # Pass empty model_config_dict so nothing is spread to create() call
         
         return ModelFactory.create(
             model_platform=ModelPlatformType.OPENAI,
             model_type=llm_model,
-            model_config_dict=model_config,
+            model_config_dict={},
         )
     
     def _get_active_agents_for_round(