feat(benchmark): expose logical worker endpoints

[cquil11]

Summary

• Inject mode-specific logical worker leader IPs and endpoints into custom benchmark processes.
• Give benchmark.type: custom an ordered AIPERF_SERVER_METRICS_URLS without requiring Slurm hostlist parsing or duplicated topology metadata.
• Preserve the existing physical-process metrics behavior for built-in AIPerf runners, including vLLM data-parallel layouts.
• Document the environment contract and test disaggregated, aggregated, multinode, co-located, and built-in-AIPerf cases.

Motivation

srt-slurm already owns endpoint allocation, node placement, port assignment, heterogeneous-job topology, and network-interface-aware IP resolution. A custom benchmark currently cannot consume that authoritative runtime mapping. Consumers must reconstruct it from SLURM_JOB_NODELIST, recipe-specific node counts, and the system-port allocation algorithm.

That reconstruction is brittle and cannot distinguish a logical worker leader from a multinode follower rank. InferenceX PR #2291 is a concrete consumer: its custom AgentX/AIPerf replay needs SGLang worker metrics to drain requests correctly between concurrency points while running against an unmodified upstream srt-slurm release.

Environment contract

When a mode is present, custom benchmark processes receive:

Mode	IPs	Endpoints
Prefill	SRT_PREFILL_IPS	SRT_PREFILL_ENDPOINTS
Decode	SRT_DECODE_IPS	SRT_DECODE_ENDPOINTS
Aggregated	SRT_AGG_IPS	SRT_AGG_ENDPOINTS

• IP and endpoint lists are comma-separated and preserve logical topology/index order.
• Endpoints use IP:port format.
• Dynamo uses each logical leader's DYN_SYSTEM_PORT; other frontends use the worker HTTP port.
• Multinode follower ranks are excluded.
• Co-located logical workers retain repeated IPs and distinct ports, keeping list positions aligned.
• Custom benchmarks also receive AIPERF_SERVER_METRICS_URLS as fully qualified /metrics URLs in the same order. Configured KVBM metrics URLs remain appended after worker URLs.
• benchmark.env is applied last, so recipes can override any generated value explicitly.

This contract is computed from backend_processes, the same cached topology used to launch workers. It does not hardcode SGLang recipe layouts or Slurm hostlist rules and therefore also covers aggregated/disaggregated layouts across upstream backends.

Compatibility

This is additive for benchmark environments and requires no recipe schema change. Built-in AIPerf runners retain their historical physical-process URL list because vLLM data-parallel layouts rely on it; only custom commands use the new logical-worker AIPerf view.

Validation

• uv run pytest tests/test_benchmarks.py::TestCustomBenchmarkRunner tests/test_node_placement.py -q — 15 passed.
• uv run ruff check src/srtctl/ — passed.
• uv run ruff format --check src/srtctl/ — passed.
• Full local suite: 883 passed, 2 skipped, 5 failed. All five failures reproduce unchanged on clean main at 38723f6 and are macOS/environment baseline failures (sched_getaffinity, CPU allocation metadata, and the existing SA-Bench profiling-script path test).
• DCO sign-off is present on the commit.