Which is still object-granularity memory tagging, and under no circumstances it can prevent a program from using its own memory any way it wants (e.g. by simply never requesting more than one object), and they obviously don't even claim to do so.

In order to create a system where you can't have a way to avoid bounds checking (as the above post was claiming), you would basically have to throw turing-completeness away, e.g. by preventing the implementation of arrays, lists, etc. The moment you can implement practically any data structure whatsoever, you can use it to build an allocator on top of it (through a heap... or not), and then proceed to access each (sub)object allocated by it without any bounds checking whatsoever. From the point of view of the hardware/OS, you will only have one large object. This is literally how most user programs work today (from the point of view of the hardware/OS).

You can provide a system allocator that tags objects separately as you allocate, but there is no way to prevent a user program from managing its own memory as it sees fit.