feat: add vllm sidecar auto-registration.

[fdvty]

Description

Adds a standalone vLLM sidecar (vllm_sidecar/) that auto-registers a vLLM
instance into xllm-service's etcd, replacing the manual register_vllm.sh step
from the previous PR. Components:

• etcd_registry.py — etcd gateway client (put/lease/keepalive for the instance key)
• health.py — polls the vLLM /health endpoint before registering
• meta.py — builds the InstanceMetaInfo JSON (backend_type=vllm, type=DEFAULT)
• sidecar.py — the runtime entrypoint tying them together
• tests for meta and the sidecar runtime

Related Issues

Change Type

• New feature
• Bug fix / Perf / Refactor / Docs / Test / Build or CI

Pull Request Checklist

• Commit/PR title follows <type>: <subject>.
• Rebased onto the latest main.
• No C++ changed; Python formatted (ruff) and unit-tested (15 cases pass).

Reviewer Notes

• Self-contained new Python package; ~1/3 of the diff is unit tests.
• Builds on the vLLM backend merged in feat: connect xllm-service http entrypoint to vllm backend. #68; supersedes the manual
  register_vllm.sh registration with automatic etcd registration.

[gemini-code-assist]

Code Review

This pull request introduces a new vllm_sidecar package to automatically register and maintain the liveness of a vLLM instance in the xllm-service cluster using etcd leases. The feedback focuses on enhancing the robustness of the HTTP-based etcd client and the health prober. Key recommendations include handling pre-existing schemes in etcd endpoints, safely parsing JSON responses to prevent crashes from invalid or null fields, and using a requests.Session in the health prober to enable connection reuse and avoid socket exhaustion.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[fdvty]

Thanks for the thorough review! Addressed every comment, and proactively audited the whole PR for the same classes of issues:

• etcd_registry.py — hardened the entire JSON-parsing surface against proto3 omit-empty/null: non-object 200 bodies coerce to {}, non-JSON 200 falls through to the next endpoint, lease_keepalive tolerates null result/TTL, get tolerates an omitted value, and endpoint scheme prefixes are preserved.
• health.py — VllmHealthProbe now reuses a single requests.Session (keep-alive) instead of a socket per poll.
• sidecar.py / meta.py / demo — fixed a stale doc comment and removed a heartbeat-only flag (--xllm-service-url) from the demo startup that this PR's sidecar doesn't accept yet. Added unit tests for each new edge case (21 tests total). Suggest Squash and merge to collapse the review fixes into the feature commit.