feat(inference): implement run_episode() — full reset→step loop with httpx client

inference.py

@@ -39,3 +39,25 @@ def generate_attack(defender_response: str, turn: int, previous_success: float)
         "intensity":       min(0.3 + (turn * 0.1), 1.0),
         "framing":         framing,
     }
+
+async def run_episode(task: str = "easy") -> dict:
+    import httpx
+    async with httpx.AsyncClient(base_url=API_BASE_URL, timeout=60.0) as http:
+        resp = await http.post("/reset")
+        reset_data = resp.json()
+        defender_resp = reset_data["observation"]["defender_response"]
+        prev_success = 0.0
+        
+        turn = 0
+        while True:
+            turn += 1
+            action = generate_attack(defender_resp, turn, prev_success)
+            resp = await http.post("/step", json=action)
+            step_data = resp.json()
+            obs = step_data["observation"]
+            defender_resp = obs["defender_response"]
+            prev_success = obs["attack_success_estimate"]
+            if obs["episode_done"]: break
+        
+        grade_resp = await http.post("/grade")
+        return {"task": task, "turns": turn, "grade": grade_resp.json()}