the latency was never the network, it was the model thinking. every reply came back with four hundred reasoning tokens and nothing else, the phone loop spent its whole budget on internal monologue and returned empty. turn reasoning off and the same calls go from five seconds to an average of four hundred milliseconds, with zero empty replies. switching the provider sort to throughput shaved it further. the whole fix was one flag and an ordering preference, not a bigger model.
the timer routine failed six times before it passed. three separate causes. the element list was cut off at forty five and the clock keypad digits live at indexes forty to fifty one, so the planner could not see the keys it needed. indexes are re issued on every screen, so a saved index meant delete, and it deleted the timer it just made. the fix was to stop using a model for a fixed goal and compile the routine, one gesture per step, every lookup fresh, end state verified. it passes in about twenty nine seconds now.
then i tested it through the app the way a person actually uses it, typed into the command box, and four real bugs fell out. the result line never updated because the screen only read the stored value on resume. raw json printed on screen instead of a human line. the app could not reach the runner from the cloud box. it opened a gated music app. fixed each one: poll the result every second, render done with the real time, try the tunnel addresses in order, map the gated apps to the open one only when it is installed.