fix: Simpler prompt format and temperature=0 for Devstral

- Remove newlines from instruction (they become <SPECIAL_*> tokens)
- Add explicit <s> BOS token to manual format
- Use temperature=0 for fully deterministic output
- Clearer instruction: "Continue this Python code:"

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <[email protected]>

backend/model_service.py

@@ -1510,12 +1510,12 @@ async def analyze_research_attention(request: Dict[str, Any], authenticated: boo
                 logger.info(f"Applied chat template for {manager.model_id}")
             else:
                 # Fallback: Manual Mistral-style instruction format
-                formatted_prompt = f"[INST] Complete the following code:\n{prompt} [/INST]"
+                # Keep it simple - no newlines (they become special tokens)
+                formatted_prompt = f"<s>[INST] Continue this Python code: {prompt} [/INST]"
                 logger.info(f"Applied manual instruction format for {manager.model_id}")
-            # Use lower temperature for instruct models (more deterministic code)
-            if temperature > 0.3:
-                temperature = 0.15
-                logger.info(f"Adjusted temperature to {temperature} for instruct model")
+            # Use temperature=0 for instruct models (fully deterministic code)
+            temperature = 0.0
+            logger.info(f"Using temperature={temperature} for deterministic instruct model output")
         else:
             # Base model - use raw prompt
             formatted_prompt = prompt