[E2E][Deep Agents] Standard repo install fails only /v1/models despite successful chat

## Description

The standard repository-install E2E target for LangChain Deep Agents Code completes onboarding and successfully runs real managed inference, but the final headless-inference check fails because `GET https://inference.local/v1/models` returns HTTP 503.

This is not a sandbox creation or ownership failure: the OpenShell gateway starts, the sandbox reaches `Ready`, the create ownership handoff completes, registry/state validation passes, and the DCode runtime is ready.

## Exact failing run

- Workflow run: https://github.com/NVIDIA/NemoClaw/actions/runs/33185369934
- Job: https://github.com/NVIDIA/NemoClaw/actions/runs/33185369934/job/98898906810
- Main SHA: `b6ee5c8caf2824a945ca3ed3b7ed3e92088982af`
- Target: `ubuntu-repo-cloud-langchain-deepagents-code`
- Agent: LangChain Deep Agents Code v0.1.55
- Provider/model: compatible endpoint, `nvidia/nvidia/nemotron-3-ultra`
- OpenShell: 0.0.106

## Actual result

The headless-inference script reports 17 passed checks and 1 failure:

```text
OK (login-shell dcode -n reached managed inference with json-pong (exit 0; direct DNS/hosts absent))
OK (direct-exec dcode -n reached managed inference with json-pong (exit 0; direct DNS/hosts absent))
OK (nemoclaw connect --probe-only accepted the managed inference route (direct DNS/hosts absent))
FAIL: login-shell proxy did not receive HTTP 200 from https://inference.local/v1/models (HTTP 503)
```

The evidence narrows the failure to model discovery through the managed proxy. Both real Chat Completions calls succeed through the same managed route, so this is not a general inference outage or invalid credential.

## Expected result

After successful standard install and real managed inference, `GET /v1/models` should return a successful model-discovery response and the exact target should pass.

## Acceptance criteria

- Reproduce and classify why `GET /v1/models` returns 503 while Chat Completions succeeds through the same managed route.
- Preserve a bounded, credential-safe diagnostic for the 503 response (status plus sanitized failure classification/body).
- Fix the readiness, routing, or upstream model-discovery behavior; do not make the target green by skipping the real probe.
- Keep the successful login-shell/direct DCode inference and `connect --probe-only` assertions.
- Pass `ubuntu-repo-cloud-langchain-deepagents-code` on the exact tested head.

## Related, not duplicate

- #10551 also reports an Ultra `/v1/models` 503, but its real DCode chat request fails. In this E2E reproduction, real chat succeeds twice and only model discovery fails.

## Checklist

- [x] Reproduced in CI
- [x] Searched existing issues
- [x] No credentials or secrets included