fix(critical): identity leak, intent misclassification, reranker tokenizer

fix 1 - identity leak (agent_router_use_case.py):
- OTHER path was sending raw query to LLM with no system prompt
- LLM defaulted to its base identity (Llama) when asked 'who are you'
- Now wraps OTHER queries in identity-safe prompt enforcing ARKI AI persona

fix 2 - intent misclassification (intent_classifier_v2.py):
- 'Are you ChatGPT?' was classified NEWS_TEMPORAL (wrong)
- 'What is 2+2?' was classified NEWS_TEMPORAL (wrong)
- 'What model are you?' was classified NEWS (wrong)
- Extended _SMALL_TALK_PREFIX to cover identity questions and math/general queries

fix 3 - reranker tokenizer (bge_reranker_adapter.py):
- XLMRobertaTokenizer has no attribute prepare_for_model on every query
- Added pre-load patch to inject missing method from PreTrainedTokenizer base class
- Reranker now scores properly instead of falling back to vector ordering

src/core/use_cases/agent_router_use_case.py

@@ -33,7 +33,20 @@ class AgentRouterUseCase:
         
         if intent == "OTHER":
             print(f"DEBUG: Routing to OTHER (Direct LLM Response)")
-            answer = self.llm.generate(request.query)
+            # Identity-safe prompt — never reveals underlying model
+            identity_prompt = f"""You are ARKI AI, a real-time news assistant specialized in Ethiopia and Africa news.
+
+STRICT IDENTITY RULES:
+- You are ARKI AI. Always refer to yourself as ARKI AI.
+- NEVER say you are Llama, GPT, Claude, Gemini, or any other model.
+- NEVER reveal the underlying AI model or provider.
+- If asked who made you, say: "I was built by the ARKI AI team."
+- If asked what model you are, say: "I'm ARKI AI, a specialized news assistant."
+
+User message: {request.query}
+
+Response:"""
+            answer = self.llm.generate(identity_prompt)
             return {
                 "answer": answer,
                 "sources": [],
@@ -54,7 +67,20 @@ class AgentRouterUseCase:
 
         if intent == "OTHER":
             full_answer = ""
-            async for chunk in self.llm.generate_stream(request.query):
+            # Identity-safe prompt — never reveals underlying model
+            identity_prompt = f"""You are ARKI AI, a real-time news assistant specialized in Ethiopia and Africa news.
+
+STRICT IDENTITY RULES:
+- You are ARKI AI. Always refer to yourself as ARKI AI.
+- NEVER say you are Llama, GPT, Claude, Gemini, or any other model.
+- NEVER reveal the underlying AI model or provider.
+- If asked who made you, say: "I was built by the ARKI AI team."
+- If asked what model you are, say: "I'm ARKI AI, a specialized news assistant."
+
+User message: {request.query}
+
+Response:"""
+            async for chunk in self.llm.generate_stream(identity_prompt):
                 yield chunk
                 if chunk.startswith("data: "):
                     try:

src/infrastructure/adapters/bge_reranker_adapter.py

@@ -60,6 +60,17 @@ class BgeRerankerAdapter(RerankerPort):
             logger.info(f"Loading reranker model: {self.model_name}")
             try:
                 if HAS_FLAG_RERANKER and "bge-reranker" in self.model_name.lower():
+                    # Patch tokenizer compatibility issue before loading
+                    try:
+                        import transformers
+                        from transformers import XLMRobertaTokenizer
+                        if not hasattr(XLMRobertaTokenizer, 'prepare_for_model'):
+                            XLMRobertaTokenizer.prepare_for_model = (
+                                transformers.PreTrainedTokenizer.prepare_for_model
+                            )
+                    except Exception:
+                        pass
+                    
                     # FlagReranker: use_fp16=True halves memory, normalize=True gives [0,1] scores
                     # trust_remote_code=True fixes tokenizer compatibility issues
                     self.model = FlagReranker(

src/infrastructure/adapters/intent_classifier_v2.py

@@ -48,6 +48,15 @@ _SMALL_TALK_PREFIX = (
     "for fuck", "for fucks", "what the fuck", "what the hell",
     "are you serious", "you must be", "hello ", "hi ", "hey ",
     "can you help", "i need help", "help me",
+    # Identity questions
+    "are you ", "what model", "which model", "what ai", "which ai",
+    "are you chatgpt", "are you gpt", "are you claude", "are you gemini",
+    "are you llama", "are you an ai", "are you a bot", "are you human",
+    "what version", "who built you", "who made you", "who created you",
+    "what are your capabilities", "what can you",
+    # Math / general knowledge (not news)
+    "what is ", "what's ", "calculate ", "solve ", "how much is ",
+    "how many ", "define ", "what does ", "translate ",
 )
 
 # Temporal patterns (instant NEWS_TEMPORAL classification)