feat(vllm): route RL controls through sidecar

[connorcarpenter15]

Summary

• Route Dynamo vLLM sidecar RL controls and updates through the companion native vllm.Control gRPC API instead of a development HTTP admin server.
• Advertise only server-reported pause/resume, sleep/wake, weight-transfer, draft-update, and weight-version capabilities; preserve vLLM's RL HTTP request schemas at the Dynamo /engine/* boundary.
• Add opt-in Rust worker RL discovery at dyn://<namespace>.<component>.rl, including safe endpoint shutdown and discovery unregister/register ordering around pause, sleep, resume, and wake operations.
• Keep the generic vLLM mocker fail-closed while adding a focused sidecar fake that verifies route payloads reach the expected protobuf calls.

Dependencies and overlap

• Companion vLLM implementation: connorcarpenter15/vllm#22, pinned in the vendored protocol README at commit 7f3ab290464ac319e867b0d011d11dd6b2ff37f4.
• This branch is based on current main and includes the four commits from ai-dynamo/dynamo#12734 because the split native Control/Inference protocol is a prerequisite. Once that PR merges, GitHub should reduce this PR to the RL-specific diff.
• ai-dynamo/dynamo#11804 also explores RL discovery but publishes a vLLM HTTP admin URL. This PR is materially different: it keeps administration on the native gRPC sidecar channel, capability-gates every route, and publishes only Dynamo engine routes plus the existing system URL.

Validation

• cargo fmt --all -- --check — passed.
• cargo check -p dynamo-vllm-sidecar -p dynamo-backend-common -p dynamo-vllm-mocker -p dynamo-trtllm-sidecar -p dynamo-mocker-backend — passed.
• cargo check --manifest-path lib/bindings/python/Cargo.toml — passed.
• cargo clippy -p dynamo-vllm-sidecar -p dynamo-backend-common -p dynamo-vllm-mocker -p dynamo-trtllm-sidecar -p dynamo-mocker-backend --all-targets -- -D warnings — passed.
• cargo test -p dynamo-vllm-sidecar rl_engine_routes_preserve_lifecycle_payloads_and_version -- --nocapture — passed, 1 test.
• cargo test -p dynamo-vllm-sidecar aggregated_generation_converts_request_stream_and_usage -- --nocapture — passed, 1 test.
• cargo test -p dynamo-backend-common routes_request_describes_the_worker_engine_surface -- --nocapture — passed, 1 test.
• cargo test -p dynamo-backend-common engine_control_policy_wraps_discovery_mutating_controls -- --nocapture — passed, 1 test.

Not run

• No GPU/real-engine test was run locally. The cross-process request and lifecycle contracts are covered with the in-process tonic fake; end-to-end validation requires a vLLM build containing the companion draft PR.

AI assistance disclosure

This implementation was prepared with OpenAI Codex assistance and requires human review before merge.