It's a cute theorem, but it's more misleading than it is useful. Let's say we use a simple bytecode format for your programs. They can only open that dialog by having a call to the open-dialog function with a fixed parameter saying "VIRUS". We may not be able to say if an arbitrary program will run that call, but we can state with certainty whether it has that call. Such a program is either a virus or a virus decoy, and either way should be blocked.
We don't need to answer the question "Are we certain the program does X at runtime?". If we can answer "Does the program have an X-performing module?" that's good enough.
On android we can't answer that because the execution model is too powerful. But it's possible to make a practical app platform with the ability to answer that question.
That doesn't mean that you can't catch most of the cases where a programme would do something like this. As they say, perfect is the enemy of good—in this case perfect is provably impossible, but that doesn't mean we shouldn't strive to find something that improves the current state. Google automatically scans all Play Store submissions for known malicious behaviour, they might be able to detect rowhammer exploits the same way.