standup: probe decode vllm port

[hubunt]

Summary

Fix decode pod probe ports so they target the actual vLLM bind port when routing proxy is enabled.

Motivation

Fixes #1435.

In pd-disaggregation, decode vLLM is started with --port $VLLM_METRICS_PORT, which resolves to 8200. However, the rendered ModelService values used
decode.vllm.servicePort (8000) as the default decode probe port. This can cause decode pod readiness checks to fail even though vLLM is listening on 8200.

What changed

• Decode probe defaults now follow the vLLM bind port:
  • routing proxy enabled: decode.vllm.port
  • routing proxy disabled: decode.vllm.servicePort
• Existing explicit decode.probes.*.port overrides are preserved.
• Updated the port selection docs.
• Added a regression test rendering the real pd-disaggregation scenario and verifying decode startup/liveness/readiness probes use 8200.

Testing

• python -m pytest tests/test_decode_probe_port.py -q
• python -m pytest tests -q
• python -m py_compile tests/test_decode_probe_port.py
• git diff --check

Backward Compatibility

Scenarios with explicit probe port overrides are unchanged. When routing proxy is disabled, decode probes continue to default to decode.vllm.servicePort.