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
The decode benchmark timed the first requests after start, so it measured warm-up along with decoding.
Change
One timed warm-up request before the decode benchmark, reported on its own. Two runs per arm.
Result
Steady decode was within 2% in every arm. The graph-less arms paid about 40 seconds on their first request instead of at startup. For a caller, wake to first full answer went from about 135 seconds to about 125 at best: 7%, not 38%.
The same arms, judged three ways · Time to Ready
Switch views: a 38% shorter start shrinks to a 7% shorter wait once the first answer is on the clock.
Decision
From here on, a change was judged on wake to first answer, not on time to Ready.