Hover or tap a segment for its time. The twin's readiness probe runs every 10 s, so totals land on 10-second steps and the last segment absorbs the wait.
Problem
Tuning the start by trial on the live model would break callers and produce one noisy number per attempt.
Change
A disposable twin of the production server, launched as an Argo workflow with the same model, flags, declared GPU memory and memory limit. A harness reads each run's vLLM log and pod conditions and splits the start into named phases. Each run writes a ledger row before anyone interprets it.
Result
The rig turned up its own defects as it went. The twin was killed for memory at a 4 GiB limit, the workflow engine's artifact endpoints returned errors so the harness now reads object storage directly, a pod that never became Ready was parsed as zero seconds, and the decode benchmark read empty logs.
The baseline was 141 seconds to Ready. 104 of them were engine work around CUDA graphs: 52 seconds profiling their memory and 52 capturing them.
Decision
Later changes ran on the twin first, most of them two or three times on the same card, and were compared against this baseline.