[CI] Pass the scale-out endpoint flag in EC E2E

[khluu]

Summary

• replace the removed VLLM_ENABLE_SCALE_OUT_ENDPOINTS=1 setting in the scale-out EC E2E launcher
• pass --enable-scale-out to both ordinary vllm serve processes that need /inference/v1/generate

Why

Merged PR #55176 replaced the environment gate with the vllm serve --enable-scale-out flag, but the L4 Scale-out EC E2E script still uses the removed environment variable. Main build #88790 therefore started healthy encode and prefill servers without the generate route. Both the original job and its automatic retry failed the first encode request with HTTP 404.

The exact Scale-out EC job did not run in PR #55176's green build #88768. This script is one of the job's direct source dependencies, so this change also forces exact PR-CI coverage of the repaired invocation.

Duplicate check

This is not a duplicate. I reviewed issue #55117 and its comments, searched open PRs by the issue number and by enable-scale-out, Scale-out EC Connector E2E, and the 404 signature. The nearby open results concern Rust route support or unrelated scale-out features; none updates this E2E launcher.

Validation

• bash -n tests/entrypoints/scale_out/ec_integration/run_scale_out_ec_e2e_test.sh
• shellcheck tests/entrypoints/scale_out/ec_integration/run_scale_out_ec_e2e_test.sh
• pre-commit run --files tests/entrypoints/scale_out/ec_integration/run_scale_out_ec_e2e_test.sh
• git diff --check
• static assertion: exactly two --enable-scale-out arguments and no removed environment-variable reference remain in the script

All pass locally. The exact two-L4 E2E job is required before merge. No model evaluation was run because this changes only the CI harness invocation, not serving output or model code.

AI assistance

OpenAI Codex was used to diagnose the main-CI failure, prepare this patch, and run the listed checks. A human submitter must review every changed line and understand the endpoint-registration behavior before merge.

[khluu]

Exact hardware validation is terminal green at c3bf36938241ec1d092147fffdcd57254d32feb6.

• Build: #88826 — passed
• Exact lane: L4 Scale-out EC Connector E2E — exit 0 in 3m45s
• Evidence: baseline and scale-out services started successfully; the single-image, two-image, and text-only scale-out outputs all matched the baseline.

Lightweight checks and the exact two-L4 machine gate are now green. The remaining gate is human review.

[mergify]

Documentation preview: https://vllm--56819.org.readthedocs.build/en/56819/

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[tlrmchlsmth]

/ci run

[github-actions]

✅ Triggered Buildkite CI #88875 for commit b1452f4c1d39.

[khluu]

Fresh main evidence strengthens the gate for this fix:

• main #88913 original reached healthy render/encode/prefill services, then failed POST http://localhost:19601/inference/v1/generate with HTTP 404.
• Its automatic retry failed with the identical 404.
• daily-main #88920 then reproduced the same signature again.

The isolated exact lane on this PR remains green in #88826. This is now repeated deterministic main breakage, not a one-off runner failure.