Problem

✨ [AI-generated issue]

The issue-4462-gateway-pinned-approval-characterization-e2e nightly E2E job fails intermittently because the gateway WebSocket becomes transiently unreachable after the legacy approve characterization step deliberately provokes a failure. The subsequent recovery openclaw devices approve fires immediately and hits gateway connect failed.

Evidence

• Failing run: https://github.com/NVIDIA/NemoClaw/actions/runs/27450816965
• Job: issue-4462-gateway-pinned-approval-characterization-e2e / run (log)
• Error line: FAIL: recovery after legacy characterization: openclaw devices approve failed... gateway connect failed: G
• Classification: infra_flake
• Confidence: medium
• Last green nightly (same job): run 27386272836 (June 12)

Root Cause

Commits b747bfa and 09a5c69 hardened the recovery proxy-env sourcing path between the June 12 and June 13 nightlies. The legacy approve's deliberately failed WebSocket request can leave the gateway briefly unstable. The test's legacy_gateway_pinned_approval_characterization() function calls the recovery approve_request immediately without any backoff, so on slow gateway recovery the approve hits a closed connection.

Fix

Fixes #5375

Insert a gateway-readiness polling loop (5 attempts x 2 s) before the recovery approve call in test/e2e/test-issue-4462-scope-upgrade-approval.sh. The loop polls device_state_json (which reads device state files directly, not via WebSocket) to confirm the sandbox is reachable before issuing the approve.

Suggested Labels

bug, e2e, nightly