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.
Hypothesis
The 40 seconds looked like kernel compilation repeated on every pod start, because Triton's and FlashInfer's caches lived in the container's temporary home directory.
Change
Both caches moved to the persistent volume.
Result
Both caches were written and read on later runs, with no change in the start or the first request. vLLM warns on every Triton compile during inference, and only one kernel compiled. The engine logged nothing for about 40 seconds, then ran the warm-up request in about 5. A neighbouring pod on the same image held a 6.5 MB cache from the CUDA driver itself, also in a temporary directory.
Decision
Falsified: the 40 seconds wasn't Triton or FlashInfer compiling. The driver's own cache became the next lead.