wanted the assistant's on-device model to run while the assistant overlay sits on top of whatever app is on screen. that is the natural way to use it, so it seemed like a safe assumption. the docs said otherwise.
the platform's built in small model has a foreground rule. the calling app has to be visible and in the foreground, and there is a documented error for background use. an overlay window is not documented to satisfy that rule. the existing assistant already hit the error, which is why it hides behind a picture in picture window to keep inference alive.
the alternative is running a model inside the app process instead of borrowing the platform's. it works behind an overlay, because it is just your own process doing compute, but you pay for it in memory and cpu, and a small model still wants a couple of gigabytes of ram on the device. there is also the window question on current android: the system placed session window for assistant apps is the one that is still fully supported when the app holds the assistant role, while raw overlay windows keep losing privileges.
the lesson is to read the fine print before betting the architecture. the assumption that an overlay counts as foreground was wrong, and the research saved a rewrite. the plan is now a model inside the app process, with the platform model used only when it is actually in the foreground.