[Bugfix] Wire output buffers into duplex model test harnesses

[linyueqian]

The main CUDA build 16680 failed after #7644 was merged, leading to eight PersonaPlex test failures and four Nemotron fixture errors. This happened because both model-specific harnesses construct an OpenDuplexSessionMessage without the newly required output_buffer argument, as these files landed after the PR base and were missing from the tested PR head.

To fix this, we now create a DuplexOutputBuffer using the same runtime limits that each manager uses and pass it directly to both the open message and the shared Harness. We also switched to using keyword arguments for the Nemotron harness to avoid issues with the changed dataclass field order. This keeps the production argument required and lets existing tests drain events through the shared buffer. Only the two test files change.

Validation shows that local pre-commit passes. On the h200 host (observed NVIDIA L20X, vLLM 0.30.0 and PyTorch 2.13.0+cu129), the three targeted test files produced eight failures, four fixture errors and 52 passes on main commit f781f7bb, and then 64 passes on the repair commit 2337af01, using isolated sources and the same GPU allocation. The exact validation command was pytest -sv tests/engine/duplex/test_session_runner_personaplex.py tests/engine/duplex/test_session_runner_nemotron_voicechat.py tests/engine/duplex/test_session_runner.py -m "core_model and cpu" --tb=short. The main CUDA CI for the repair remains pending.

Failure evidence: https://buildkite.com/vllm/vllm-omni/builds/16680#01a108c1-420b-4bb4-b3ae-a7b7790bb7fe

[linyueqian]

Self-review: the changes are limited to the two model-specific test harnesses and output_buffer remains required in production. Each test fixture now passes a single buffer, configured from the manager's runtime limits, into both OpenDuplexSessionMessage and Harness.

Validation on the same h200 allocation with vLLM 0.30.0 shows main f781f7b reproduces 8 failures, 4 fixture errors, and 52 passes, while repair 2337af0 passes all 64 tests across the affected model suites and the generic runner suite. Local pre-commit passes and main CUDA build 16682 is pending.

@NolenLiang could you review this small follow-up to #7644 so we can restore main before resuming merges?

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/engine_orchestration.md.

Module owners: @fake0fan @tzhouam @NickCao

Routing: @fake0fan via semantic router; @tzhouam via semantic router; @NickCao via CODEOWNERS

@linyueqian, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[vllm-omni-review-bot]

Omni ReviewBot routing record

Assigned Strict on zcode (GLM-5.3-Flash) under experiment fleet-strict-cursor-grok46-zcode-glm53flash-5050-c5-z10-20261002.

[andyluo7]

Reviewed exact head 2337af0. The two test harnesses now create a bounded DuplexOutputBuffer from the same runtime limits and pass that exact buffer to both OpenDuplexSessionMessage and Harness, matching the generic runner contract introduced by #7644. The previously deterministic 8 failures and 4 fixture errors are resolved in both AMD #13285 and CUDA #16682: Simple · Engine&Entrypoints Test passed with 3709 passed on each platform. The NPU A5 failure is pre-start infrastructure (exit -1, no agent or log) and is unrelated to this test-only fix.