fix(vllm): drop reasoning-parser none on the Python sidecar launcher

[peilii]

Summary

Dynamo's vLLM sidecar refuses to start when vLLM runs its own reasoning parser alongside the sidecar's --dyn-*-parser flags (ai-dynamo/dynamo#15684), and Dynamo's launch scripts start vLLM with vllm-rs serve --reasoning-parser none. In srtctl that value only works for single-node sidecar endpoints, which launch vllm-rs serve. Hybrid-DP endpoints and cross-node replicas launch through Python vllm serve (VLLM_USE_RUST_FRONTEND=1), and Python's engine args look every --reasoning-parser value up in the parser registry; its disabled value is the absent flag. With one recipe value, prefill (DEP2, vllm-rs) came up and every decode node (DEP16 over four nodes, hybrid) died at engine-config time:

KeyError: "Reasoning parser 'none' not found. Available parsers: cohere_command3, ..., deepseek_v41, ..."

This change drops reasoning-parser: none (case-insensitive) on the Python launcher only. That is semantics-preserving: the Rust frontend Python bootstraps defaults to no parser (default_py_bootstrap_parser_selection returns ParserSelection::None in vllm-rs), so the result equals vllm-rs serve --reasoning-parser none. auto has no equivalent on this launcher (Python has no auto selection and the bootstrapped frontend never sees the flag), so it is rejected with a clear ValueError rather than forwarded into the same KeyError. Explicit parser names pass through unchanged.

Validation

• pytest tests/test_sidecar_backends.py: 67 passed, including three new tests (hybrid DP drops none/None and forwards a parser name; hybrid DP rejects auto; a single-node vllm-rs serve endpoint keeps --reasoning-parser none).
• ruff check / ruff format --check clean.
• Observed on oci-jhb with vllm/vllm-openai:nightly-aarch64-dev-20261012-59e75f4 (vLLM 0.31.1rc1.dev296+gfe4df3df0) and ai-dynamo 1.6.0.dev20261010: the failing 5P1D job (--reasoning-parser none on all roles) vs. the same recipe without the flag on decode, which reached ready with all nine sidecars serving. With this patch and reasoning-parser: none on every role, the same 5P1D deployment launched prefill as vllm-rs serve … --reasoning-parser none and decode through the Python launcher without the flag, hit no KeyError, and reached ready with all 9 sidecars serving (5 prefill, 4 decode) before the benchmark started.

Notes

• docs/sidecars.md gets one sentence in the multi-node DP paragraph.
• Longer term, if vllm-rs serve gains hybrid-DP startup, decode can leave the Python launcher and this translation becomes unnecessary.

[github-actions]

src/srtctl diff: +15 -0