bug(e2e): Stabilize nondeterministic fixture request assertions in device-auth-health and openclaw-inference-switch

[sandl99]

Investigation Summary

• device-auth-health and openclaw-inference-switch synchronously inspect fixture request ledgers for /v1/chat/completions traffic that the tests do not explicitly trigger or await.
• The same assertions pass and fail across commits that do not change the relevant tests or runtime behavior.
• Failed runs still show successful setup/onboarding and authenticated /v1/models traffic; the expected chat-completions entry is absent when the ledger is read.
• openclaw-inference-switch fails before exercising the switch, so the failure does not establish that inference switching is broken.
• The exact source of the asynchronous request timing remains unknown.

Description

The live E2E targets device-auth-health and openclaw-inference-switch contain nondeterministic fixture request assertions. Each test assumes an incidental /v1/chat/completions request has already reached its fake provider, then immediately reads the fixture ledger.

When startup or onboarding has only issued /v1/models probes by that point, the assertion fails even though setup succeeds. The tests should explicitly produce and await the request required by their contract, or wait for a bounded deterministic fixture condition before asserting.

Observed sequence for openclaw-inference-switch:

• Passed at 387cb08644fe030bb85146255f4b77e3c54697d2: https://github.com/NVIDIA/NemoClaw/actions/runs/30614075121
• Failed with the missing chat-completions request at docs-only commit f8fb820159c4843a19759efc9b0e28d4aa122440: https://github.com/NVIDIA/NemoClaw/actions/runs/30616641728
• Passed at unrelated E2E-calibration commit 6ff018521edba1bab91c8dd003a0ddcd0d29a8ca: https://github.com/NVIDIA/NemoClaw/actions/runs/30628616288
• Failed with the same signature on both attempts at 7eb0369e7a0bc91f6ae99fc1eaad3cd274a34c8a: https://github.com/NVIDIA/NemoClaw/actions/runs/30633803340

For device-auth-health, the assertion passed in run 30614075121 and failed in run 30628616288 with the same class of missing incidental fixture request:

• https://github.com/NVIDIA/NemoClaw/actions/runs/30614075121
• https://github.com/NVIDIA/NemoClaw/actions/runs/30628616288

Expected: fixture assertions depend on requests explicitly initiated by the test and complete deterministically.

Actual: assertions depend on asynchronous startup/onboarding traffic being present at the instant the fixture ledger is read.

Reproduction Steps

1. Dispatch the trusted E2E workflow with the device-auth-health and openclaw-inference-switch targets.
2. Repeat the targets without changing their source.
3. Inspect the failed fixture ledger.
4. Observe that setup/onboarding completes and /v1/models is recorded, while the asserted /v1/chat/completions entry is sometimes absent.

Environment

• GitHub-hosted Ubuntu E2E runners
• OpenShell 0.0.85
• OpenClaw 2026.7.1
• NemoClaw commits and run receipts listed above

Debug Output

CI logs and uploaded E2E artifacts are attached to the linked workflow runs. The relevant failure is the fixture-ledger assertion expecting /v1/chat/completions after only /v1/models requests were recorded.

Acceptance Criteria

• Both tests explicitly trigger the request they assert, or wait on a bounded fixture condition tied to the intended behavior.
• A timeout reports the requests observed and the phase that failed.
• Focused repeated runs demonstrate stable results without relying on incidental startup timing.

Checklist

• I confirmed this bug is reproducible
• I searched existing issues and this is not a duplicate