The useful question is not whether one optimizer is universally better. It is which information the loop receives, what each acquisition costs, and whether the method can make a complete decision before the operating point moves.
The measurement is part of the algorithm.
On physical hardware, an objective evaluation is not a free function call. It can be a batch of quantum shots, an optical scan, a cloud task, or minutes of instrument time. Methods that look similar in iteration count separate when their acquisition bill is made explicit.
Retained state changes the next round.
The configuration is not disposable. Carrying it forward preserves useful context between rounds. The IQM campaign validated the resulting advantage under pre-registered hypotheses.
The boundary matters as much as the win.
Where a scalar statistic destroys the structure, where timing is incompatible, or where the objective is invisible to the measured channel, the runtime should decline. Everything else must be decided by a fair race on the actual plant.