From the point of view of software, it's impossible in principle to tell whether or not the code being executed does what the user wants it to, and only what the user wants it to. Half of that is the halting problem, the other half is that "what user wants" is an General-AI-complete problem. Moreover, the software can't even tell the difference between "the user" and "a malicious third party".

In meatspace we solve this problem with rules and laws. Software, for better or worse, moves around too fast.

I gave it a thought, and decided I don't need apps to be able to obtain highly elevated privileges, root or similar. This is, indeed, dangerous, esp. regarding ADB root access (a rogue "charger" + an accidental wrong tap[1] = totally compromised device that can be only fixed by full re-flashing). I needed my own firmware that does things my way, signed with the keys I control.

So I did. Now all the "secure" apps are happy, and I still have the control over my device's behavior.

( Okay, I've cheated - I had to sanitize androidboot.verifiedbootstate when kernel initializes, because I can't control the bootloader :( )

[1] Hm, maybe password-authenticated root access is okay, though... But not a typical "tap to allow" dialog.