Avoid native server port collisions

[khluu]

What changed

• derive a high per-job server port from BUILDKITE_JOB_ID, skipping ports already in use
• require /v1/models to report the expected workload model before declaring the server healthy
• bound server and log-follower shutdown so a failed benchmark cannot hang until the Buildkite timeout
• add stdlib-only lifecycle regression tests for port selection, model-aware readiness, and cleanup escalation

Root cause

In nightly build 350, gpt_oss_120b-b200 and minimax_m3-b200 ran concurrently on dgxB200-14 with native servers using host networking and fixed port 8000. GPT-OSS became healthy at 06:12:31; three seconds later MiniMax's readiness probe accepted that server while MiniMax was still loading. MiniMax sent 512 requests for nvidia/MiniMax-M3-NVFP4 to the GPT-OSS server and received 512 HTTP 404 responses. Its EXIT cleanup then blocked on the log-follower pipeline until the two-hour job timeout.

Build: https://buildkite.com/vllm/perf-eval/builds/350

Validation

• python3 .buildkite/test_server_lifecycle.py (4/4)
• python3 .buildkite/test_generate_pipeline.py (6/6)
• python3 -m py_compile .buildkite/test_server_lifecycle.py .buildkite/generate_pipeline.py lib/parse_workload.py
• bash -n lib/server.sh lib/run.sh lib/run_vllm_bench.sh
• shellcheck lib/server.sh lib/run.sh
• MiniMax B200 workload parsing and targeted pipeline generation with BENCH_ONLY=1
• git diff --check
• Targeted B200 validation: https://buildkite.com/vllm/perf-eval/builds/357 — 512/512 successful requests, 0 failed