arena: live-web episodes get one process each -- a single hang was poisoning whole rounds

AssistantBench pages make BrowserGym's obs extraction slow and flaky (context-destroyed
retries, unmarked frames); one step-hang kill corrupted process-wide playwright state
and every later episode instant-failed as infra. supervisor --isolate spawns one task
per process, containing the blast radius. Step deadline raised to 150s for live pages.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WsbS5x2rYsMDxP2kW3qqmQ

e2e/browser-v3/arena/supervisor.py

@@ -91,6 +91,9 @@ def main() -> None:
     # Resume window: episodes recorded after this epoch count as done, so a restarted supervisor
     # keeps the finished work of the run it replaces instead of re-running all 125.
     ap.add_argument("--since", type=float, default=0.0)
+    # Live-web suites corrupt process-wide playwright state on a single hang; one task per process
+    # contains the blast radius (measured: val.5 hang -> every later episode instant-failed).
+    ap.add_argument("--isolate", action="store_true")
     args = ap.parse_args()
     args.log = args.log or f"/tmp/supervise_{args.arm}.log"
 
@@ -102,7 +105,8 @@ def main() -> None:
         if not remaining:
             break
         print(f"[supervisor] round {rnd}: {len(remaining)} tasks remaining", flush=True)
-        proc = spawn(args.arm, remaining, args)
+        batch = remaining[:1] if args.isolate else remaining
+        proc = spawn(args.arm, batch, args)
         last_size = path.stat().st_size if path.exists() else 0
         last_change = time.time()
         while proc.poll() is None: