feat: add native gRPC sidecar module launcher

[ishandhanani]

Adds an optional --sidecar <python.module> to launch a local companion process after SGLang's native gRPC listener starts. This removes the Dynamo-specific server contract while preserving SGLang-owned child supervision and lets provider packages remain optional.

How This Was Implemented

• Replaced the Dynamo-specific flag and launcher with a module contract: main(argv) receives --sglang-endpoint <local-endpoint>.
• Kept process spawn, parent-death handling, crash watchdog, bounded stop, and forced tree cleanup in SGLang.
• Starts the sidecar only after native gRPC binds; a sidecar stop failure is logged without skipping gRPC shutdown.
• Validates a non-empty selector and the existing native-gRPC-only launch constraints.

Walkthrough

Mental model

The selected module is a locally managed process, not an SGLang plugin or registry entry. SGLang supplies only its loopback native-gRPC endpoint; the provider owns its arguments and runtime behavior.

flowchart LR
  A["launch_server --sidecar package.module"] --> B["Native gRPC listener binds"]
  B --> C["SGLang Sidecar supervisor"]
  C --> D["Spawned child imports package.module"]
  D --> E["module.main(--sglang-endpoint local endpoint)"]
  E --> F["Provider connects to native gRPC"]
  C --> G["Watchdog / parent-death handling / bounded stop"]

Loading

Lifecycle and failure behavior

The module is imported only in the spawned child, preserving optional provider dependencies and avoiding parent-process import side effects. An import or main(argv) contract failure terminates the required child and is handled by the existing watchdog; on normal shutdown SGLang stops the child before shutting down native gRPC.

Boundaries

This is intentionally a same-process-namespace sidecar interface: wildcard SGLang binds are converted to loopback for the child. It does not add provider aliases, discovery, provider-specific arguments, or a registry.

Validation

• .venv/bin/python -m pytest -q test/registered/unit/server_args/test_server_args.py::TestGrpcServerArgs — 16 passed.
• uvx pre-commit run --files python/sglang/srt/entrypoints/sidecar.py python/sglang/srt/entrypoints/http_server.py python/sglang/srt/server_args.py test/registered/unit/server_args/test_server_args.py — passed.
• Live Qwen3-0.6B launch with --sidecar dynamo.sglang_sidecar.main: native gRPC discovery succeeded, Dynamo established eight gRPC connections, and served sglang-sidecar-e2e.backend.generate.

---

CI States

Latest PR Test (Base): 🚫 Run #29924027291
Latest PR Test (Extra): ❌ Run #29924026692

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[ishandhanani]

/tag-and-rerun-ci

[ishandhanani]

@connorcarpenter15 PTAL and review this PR

[ishandhanani]

Relevant CI has passed. Getting this in as it does not affect any GPU pieces at all