Set up Cursor Cloud dev environment for sgl-model-gateway

[Moersity]

Summary

Sets up and verifies a working Cursor Cloud development environment for this repo. SGLang is a GPU-first monorepo and this VM has no GPU, so the fully CPU-runnable product — sgl-model-gateway/ (the Rust HTTP/gRPC model router) — was the target for build/lint/test/run. This PR only adds AGENTS.md (no source changes); the environment itself is captured via the VM snapshot + update script.

What was set up (captured in the VM snapshot)

• System deps: libssl-dev, pkg-config, protobuf-compiler, build-essential, redis-server, python3.12-venv.
• Rust toolchain pinned to 1.90 (matches CI scripts/ci/cuda/ci_install_gateway_dependencies.sh) + nightly for rustfmt.
• Global cargo config at /usr/local/cargo/config.toml forcing CC=gcc/CXX=g++ and linker = "g++", because the system default cc/c++ is clang, which can't find libstdc++ <cstdint> (breaks esaxx-rs) and trips rust-lld with unable to find library -lstdc++.
• pre-commit in ~/.venvs/precommit for repo-wide lint.

Update script (dependency refresh only): cargo fetch for the gateway.

Verification (Rust 1.90, CPU-only)

Check	Command	Result
Build	cargo build --bin sgl-model-gateway	Finished OK
Lint	cargo clippy --all-targets --all-features -- -D warnings	exit 0
Format	cargo +nightly fmt -- --check	exit 0
Tests	cargo test --no-fail-fast	782 passed, 0 failed (16 binaries)
Repo lint	pre-commit run --all-files	all hooks Passed

End-to-end demo (hello world)

Launched the gateway in IGW mode (--enable-igw, no GPU workers needed), then exercised control plane + data plane against a lightweight mock worker:

• /liveness → OK, /health → 200
• /parse/reasoning (deepseek_r1) correctly split <think>…</think> from the answer
• /parse/function_call (json) extracted a get_weather tool call
• Registered a worker via POST /workers (health/metadata discovered), visible in /workers and /v1/models
• Routed a real POST /v1/chat/completions client → gateway → worker → client

End-to-end gateway demo transcript
CI checks summary

Notes

• policies::tree::tests::test_tree_structure_integrity_after_stress is a randomized 8-thread stress test that can flake ("Tenant … should have positive size"); it passes in isolation and in the pinned-toolchain full run.
• The Python sglang runtime and sgl-kernel require NVIDIA CUDA and are intentionally out of scope on this CPU-only VM.

_{To show artifacts inline, enable in settings.}