fix: update default emulator IP to 20.20.20.21 for official WAA

The official windowsarena/winarena image uses 20.20.20.21 for the Windows
VM inside QEMU, not 172.30.0.2. This change updates all default internal_ip
values to match the official WAA setup.

This enables fully unattended vanilla WAA benchmark runs.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: abrichr

openadapt_ml/benchmarks/cli.py

@@ -1989,15 +1989,15 @@ def capture_vm_screenshot(ip: str, output_path: Path | str = None) -> Path | Non
 
 
 def check_waa_probe(
-    ip: str, timeout: int = 5, internal_ip: str = "172.30.0.2"
+    ip: str, timeout: int = 5, internal_ip: str = "20.20.20.21"
 ) -> tuple[bool, str | None]:
     """Check if the WAA /probe endpoint is responding.
 
     Args:
         ip: Public IP address of the Azure VM
         timeout: Connection timeout in seconds
         internal_ip: Internal IP of the Windows VM inside QEMU.
-                     172.30.0.2 for dockurr/windows:latest
+                     20.20.20.21 for dockurr/windows:latest
                      20.20.20.21 for official windowsarena/winarena
 
     Returns:
@@ -2030,7 +2030,7 @@ def check_waa_probe(
 
 
 def poll_waa_probe(
-    ip: str, max_attempts: int = 30, interval: int = 20, internal_ip: str = "172.30.0.2"
+    ip: str, max_attempts: int = 30, interval: int = 20, internal_ip: str = "20.20.20.21"
 ) -> bool:
     """Poll the WAA /probe endpoint until it responds or timeout.
 
@@ -2377,7 +2377,7 @@ def launch_benchmark_viewer(
     vm_ip: str,
     port: int = 8765,
     open_browser: bool = True,
-    internal_ip: str = "172.30.0.2",
+    internal_ip: str = "20.20.20.21",
 ) -> None:
     """Launch the benchmark viewer for monitoring a running WAA benchmark.
 
@@ -2446,7 +2446,7 @@ def cmd_viewer(args: argparse.Namespace) -> None:
     vm_ip = args.vm_ip
     port = getattr(args, "port", 8765)
     no_open = getattr(args, "no_open", False)
-    internal_ip = getattr(args, "internal_ip", "172.30.0.2")
+    internal_ip = getattr(args, "internal_ip", "20.20.20.21")
 
     launch_benchmark_viewer(
         vm_ip=vm_ip, port=port, open_browser=not no_open, internal_ip=internal_ip
@@ -3821,8 +3821,8 @@ def cmd_vm(args: argparse.Namespace) -> None:
 
         print(f"  VM IP: {ip}")
 
-        # Use 172.30.0.2 for vanilla WAA (dockurr/windows base, used by windowsarena/winarena)
-        internal_ip = getattr(args, "internal_ip", "172.30.0.2")
+        # Use 20.20.20.21 for vanilla WAA (dockurr/windows base, used by windowsarena/winarena)
+        internal_ip = getattr(args, "internal_ip", "20.20.20.21")
 
         if getattr(args, "wait", False):
             # Polling mode - keep checking until ready
@@ -4351,7 +4351,7 @@ def delete_vm(name: str) -> tuple[str, bool, str]:
         print("\n2. CURRENT ACTIVITY")
         print("-" * 70)
         if not use_mock:
-            activity = detect_vm_activity(ip, "azureuser", "winarena", "172.30.0.2")
+            activity = detect_vm_activity(ip, "azureuser", "winarena", "20.20.20.21")
         activity_icon = "⚙" if activity.is_active else "💤"
         print(f"  Status:     {activity_icon} {activity.activity_type.upper()}")
         print(f"  Details:    {activity.description}")
@@ -4530,9 +4530,9 @@ def start_server():
                 # Update status every update_interval seconds
                 if (current_time - last_update).total_seconds() >= update_interval:
                     # Quick status check
-                    is_ready, probe_msg = check_waa_probe(ip, internal_ip="172.30.0.2")
+                    is_ready, probe_msg = check_waa_probe(ip, internal_ip="20.20.20.21")
                     activity = detect_vm_activity(
-                        ip, "azureuser", "winarena", "172.30.0.2"
+                        ip, "azureuser", "winarena", "20.20.20.21"
                     )
                     status_line = f"WAA: {'READY' if is_ready else 'waiting'} | Activity: {activity.activity_type}"
 
@@ -4550,7 +4550,7 @@ def start_server():
                     last_update = current_time
                 else:
                     # Use cached status
-                    is_ready, _ = check_waa_probe(ip, internal_ip="172.30.0.2")
+                    is_ready, _ = check_waa_probe(ip, internal_ip="20.20.20.21")
                     status_line = f"WAA: {'READY' if is_ready else 'waiting'}"
 
                 # Live status display
@@ -4730,7 +4730,7 @@ def start_server():
 
         for i in range(12):
             time.sleep(10)
-            is_ready, response = check_waa_probe(ip, internal_ip="172.30.0.2")
+            is_ready, response = check_waa_probe(ip, internal_ip="20.20.20.21")
             if is_ready:
                 print("\n WAA server is running!")
                 print(f"  Response: {response}")
@@ -4943,7 +4943,7 @@ def run_diag_cmd(cmd: str) -> tuple[bool, str, str]:
         # WAA probe status
         print("\n[Bonus] WAA Probe Status")
         print("-" * 50)
-        is_ready, response = check_waa_probe(ip, internal_ip="172.30.0.2")
+        is_ready, response = check_waa_probe(ip, internal_ip="20.20.20.21")
         if is_ready:
             print(f"  ✓ WAA server READY: {response[:100] if response else '(empty)'}")
         else:
@@ -5875,7 +5875,7 @@ def step(msg: str) -> None:
     max_attempts = (max_wait_minutes * 60) // poll_interval
 
     for attempt in range(max_attempts):
-        is_ready, response = check_waa_probe(ip, timeout=5, internal_ip="172.30.0.2")
+        is_ready, response = check_waa_probe(ip, timeout=5, internal_ip="20.20.20.21")
         if is_ready:
             print(f"\n  WAA server is ready!")
             break
@@ -6458,8 +6458,8 @@ def main() -> None:
     )
     p_vm.add_argument(
         "--internal-ip",
-        default="172.30.0.2",
-        help="Internal IP of Windows VM (172.30.0.2 for vanilla WAA)",
+        default="20.20.20.21",
+        help="Internal IP of Windows VM (20.20.20.21 for vanilla WAA)",
     )
     p_vm.add_argument(
         "--yes", "-y", action="store_true", help="Skip confirmation prompts"
@@ -6570,8 +6570,8 @@ def main() -> None:
     )
     p_viewer.add_argument(
         "--internal-ip",
-        default="172.30.0.2",
-        help="Internal IP of Windows VM (default: 172.30.0.2)",
+        default="20.20.20.21",
+        help="Internal IP of Windows VM (default: 20.20.20.21)",
     )
 
     # View benchmark results - generate and serve HTML viewer for collected benchmark data