Add vLLM server intropsection endpoints

[hickeyma]

Adds to the repo the vLLM introspection endpoints which are ported from the draft at hickeyma/vllm-server-introspection.

Background is in llm-d/llm-d#2088. The ask is for vLLM to provide endpoints that llm-d could use for routing and scheduling. vLLM community pushed back on growing the core API. So the endpoints ship as endpoint plugins instead which are installed next to the server and loaded at startup. No fork, no patch and no vLLM core changes.

The endpoints

Three of them each a separate plugin so you can turn on only what you need:

• GET /plugins/llm-d-server-introspection/config: how the server was launched. Needs no engine, so it works on the CPU only render server
• GET /plugins/llm-d-server-introspection/devices: per-rank hardware. Needs the worker extension via --worker-extension-cls
• GET /plugins/llm-d-server-introspection/kv-cache: KV cache capacity and attention group structure, post-profiling

They stay off unless you name them in VLLM_PLUGINS, so installing the package doesn't change how an existing server behaves.

Layout

Top level is inference server agnostic, as agreed on the issue. vLLM code sits under vllm/ as one installable package (llm-d-api-extensions-vllm) with each extension family as a subpackage. If SGLang grows a similar hook later, it lands as a sibling directory.

What got renamed

Everything picked up llm-d naming on the way over:

	Was	Now
Package	vllm_server_introspection	llm_d_api_extensions_vllm.server_introspection
VLLM_PLUGINS names	vllm_server_introspection_*	llm_d_server_introspection_*
Routes	/plugins/vllm-server-introspection/*	/plugins/llm-d-server-introspection/*

Dropping vllm from a path that vLLM itself serves seemed like the right call and it namespaces llm-d against anyone else's plugins.

Community files

Adds standard files like CONTRIBUTING.md, MAINTAINERS.md, OWNERS, SECURITY.md and a Python .gitignore. Also CODE_OF_CONDUCT.md, CODEOWNERS, issue and PR templates, dependabot config, a Makefile and two workflows (ruff, unit tests). This follows what llm-d-benchmark and llm-d-kv-cache-manager already do.

Closes llm-d/llm-d#2088
Closes #1

[hickeyma]

@elevran @vMaroon Do you mind reviewing when you get a chance?

[elevran]

Solid, well-tested first PR for this repo. The endpoint-plugin architecture matches the stated design, naming was renamed consistently everywhere (no leftover vllm_server_introspection references), and CI passes. Found one startup-crash risk that contradicts its own docstring, a secrets-exposure gap the README already warns about, and a CODEOWNERS entry GitHub can't validate. A complexity pass also found ~180 lines of avoidable duplication (logger boilerplate x4, a response-handler pattern duplicated between devices_plugin.py and kv_cache_plugin.py, and repetitive per-kind test builders in test_kv_cache_plugin.py / test_schemas.py) - not blocking, just worth a follow-up cleanup pass.

[hickeyma]

Thanks @elevran for the review. I'll do the duplication cleanup in a follow up PR.