What happened:

The hermetic integration tests under test/integration/epp/ flake intermittently on CI with Server failed to bind port 127.0.0.1:<port> errors, causing lint-and-test to fail on unrelated PRs.

Observed in at least:

• PR ci: add PR smoke test against llm-d-kv-cache main #1058: TestFullDuplexStreamed_KubeInferenceObjectiveRequest/Standalone-NoCRD/no_backend_pods_available#01
  https://github.com/llm-d/llm-d-inference-scheduler/actions/runs/25718109453/job/75516805509
• PR update: pull GIE CRD from upstream instead of locally #1043: TestFullDuplexStreamed_GRPC_KubeInferenceObjectiveRequest/do_not_shed_requests_by_default
  https://github.com/llm-d/llm-d-inference-scheduler/actions/runs/25633770069/job/75241972222

Failure signature (different sub-tests, same root cause):

Error:    Condition never satisfied
Messages: Server failed to bind port 127.0.0.1:<random-port>

What you expected to happen:

Hermetic integration tests should pass deterministically on CI without flaking due to port allocation races.

How to reproduce it (as minimally and precisely as possible):

Flaky — run lint-and-test repeatedly on a PR; intermittently a sub-test of TestFullDuplexStreamed_* fails with the bind-port error within its 10s budget.

Anything else we need to know?:

Likely cause: the random-port allocation in the hermetic test fixture races with another process / parallel sub-test on the runner; the 10s bind retry budget expires before the port frees up.

Suggested follow-up:

• Audit the port-allocation helper in the hermetic harness; prefer a listen-then-pass-fd pattern over allocate-then-bind to avoid the race.
• Or: increase the retry/timeout budget on bind failures.

Environment:

• CI: GitHub Actions, ubuntu-latest runner (default lint-and-test workflow).
• Not environment-specific to a user setup; reproduces on CI only.