Warm dependent model servers before starting their callers

When one service performs startup warmup against another local model server, start the dependency first and wait on its health endpoint before launching the caller. A process being alive is not enough: first model loading or compilation can exceed the caller's ordinary inference timeout and make readiness fail intermittently. Use a bounded dependency-health wait, then start the caller, and keep the live request timeout unchanged so warmup does not weaken production latency limits.

Applies when: multiple inference services share a host and one calls another during startup or readiness. Keywords: model warmup startup ordering readiness timeout sidecar inference health