feat: implement multi-provider model fallback to Mistral Large 2 in cycles

backend.py

@@ -1970,20 +1970,38 @@ async def execute_build_cycle(project_name: str, goal: str):
         "Write ONE JSON instruction for Space 3 (The Forge) to code the next milestone. "
         "Respond ONLY with JSON: {\"prompt\": \"task description\", \"context_rules\": \"rules\"}"
     )
+    raw = ""
     try:
         res = await nim_client.chat.completions.create(
-            model="nvidia/llama-3.1-nemotron-70b-instruct",
+            model=RECOMMENDED_MODEL,
             messages=[{"role": "user", "content": plan_prompt}],
             max_tokens=300
         )
         raw = res.choices[0].message.content.strip()
+    except Exception as e:
+        log_activity(f"[Model Fallback] NIM failed ({e}). Retrying with Mistral Large...")
+        try:
+            if mistral_client:
+                res = await mistral_client.chat.completions.create(
+                    model="mistralai/mistral-large-2-instruct",
+                    messages=[{"role": "user", "content": plan_prompt}],
+                    max_tokens=300
+                )
+                raw = res.choices[0].message.content.strip()
+            else:
+                raise Exception("Mistral client not initialized")
+        except Exception as e2:
+            log_activity(f"[Model Fallback Error] Mistral failed too: {e2}")
+            return
+
+    try:
         if raw.startswith("```"):
             raw = raw.split("```")[1].lstrip("json").strip()
         task = json.loads(raw)
         task_prompt = task.get("prompt", "")
         context_rules = task.get("context_rules", "")
     except Exception as e:
-        log_activity(f"[Build Mode Error] NIM planning failed for project '{project_name}': {e}")
+        log_activity(f"[Build Mode Error] Plan JSON parsing failed: {e} | Raw: {raw}")
         return
 
     # Use the context-locked Bell Curve prompt builder (enforces max 2500 tokens)
@@ -2088,20 +2106,38 @@ async def execute_eternity_cycle(project_name: str, goal: str):
         "Based on this research, plan the SINGLE most impactful next feature to implement. "
         "Respond ONLY with JSON: {\"prompt\": \"task\", \"context_rules\": \"rules\"}"
     )
+    raw = ""
     try:
         res = await nim_client.chat.completions.create(
-            model="nvidia/llama-3.1-nemotron-70b-instruct",
+            model=RECOMMENDED_MODEL,
             messages=[{"role": "user", "content": plan_prompt}],
             max_tokens=300
         )
         raw = res.choices[0].message.content.strip()
+    except Exception as e:
+        log_activity(f"[Model Fallback] NIM failed ({e}). Retrying with Mistral Large...")
+        try:
+            if mistral_client:
+                res = await mistral_client.chat.completions.create(
+                    model="mistralai/mistral-large-2-instruct",
+                    messages=[{"role": "user", "content": plan_prompt}],
+                    max_tokens=300
+                )
+                raw = res.choices[0].message.content.strip()
+            else:
+                raise Exception("Mistral client not initialized")
+        except Exception as e2:
+            log_activity(f"[Model Fallback Error] Mistral failed too: {e2}")
+            return
+
+    try:
         if raw.startswith("```"):
             raw = raw.split("```")[1].lstrip("json").strip()
         task = json.loads(raw)
         task_prompt = task.get("prompt", "")
         context_rules = task.get("context_rules", "")
     except Exception as e:
-        log_activity(f"[Eternity Mode Error] NIM planning failed for '{project_name}': {e}")
+        log_activity(f"[Eternity Mode Error] Plan JSON parsing failed: {e} | Raw: {raw}")
         return
 
     # Step 3: Build Karpathy context-locked prompt (Apex)