Adds `vllm-multiinstance`: benchmarks a vLLM CPU image on AMD EPYC by
running N vLLM instances (each pinned to 32 physical cores) behind an
NGINX load balancer, driving load with guidellm via ansible, and
reporting peak aggregate memory (podman stats) + end-to-end
throughput/latency across models, concurrency rates, and instance
counts. The benchmark harness is vendored with the skill; only podman +
ansible are required.

Robust on leaner hosts: a host preflight (check-host.sh) fails fast with
actionable guidance on unresolvable image short-names, missing rootless
cgroup cpuset delegation, and CNI cniVersion mismatch; start.sh
auto-downgrades the CNI conflist, guards static-IP fallback, and
fast-fails dead containers instead of hanging the health wait;
run_sweep.sh auto-detects missing passwordless sudo and runs
ansible/guidellm rootless (ansible_become=false).

Contents: SKILL.md, skill-card.md, README.md, reference.md, vendored
harness/, scripts/; registered in .claude-plugin/marketplace.json
(+ regenerated .cursor-plugin manifest); behavioral eval at
eval/behavioral/tests/test_vllm_multiinstance.py.

Testing: structural gate (check.sh) passes with 0 errors; behavioral
eval (LLM-judged, sonnet) 5/5 passed -- covers sweep sizing, the
guidellm.log-vs-benchmarks.json score rule, host-preflight fail-fast,
image short-name remediation, and rootless/become auto-detection.

Signed-off-by: Rahul Jauhari <rahul.jauhari@amd.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Change-Id: Ifa419ea79793fcfb303b6f1cc657539b22622f8b