the phone agent's loop was flailing on the live ui, so i hardened it. the screen diff now reports added and removed texts instead of a vague changed. the element list gets deduped by screen position and sorted by score before the model reads it, so it stops seeing the same button five times and tapping the wrong one.
the stuck detection grew a ladder. two steps with no change gives an ocr hint, three tells the model its plan is not working and to form a new one, four forces a back, five forces home. the same action repeated three times gets hard refused for two steps, and the result of the last action feeds back every step so the model can learn that a tap failed instead of assuming it landed. there is a step budget now too.
then i gave a false green and had to correct it. the first run claimed all emulator runs passed, and the second run with a longer read window showed the music playback routine could never pass. the emulator runs without audio, so the playback check can never confirm, and the fake music app deliberately puts an artist card first, which the routine correctly refuses to tap blindly. the music path was never green, on the old version either. the unit tests pass and the tap routines pass, but real proof is talking to it on an actual phone.
the lesson underneath: a test window that is too short can read the success lines from the previous run. verify the verification.