The rationale: without this information, it is impossible for anybody else to write a bytecode compiler / assembler, because when you create a code object you have to specify its stack depth. I used this information for my "maynard" bytecode assembler / disassembler.
That said, I'm not sure who needs these super-fancy versions Serhiy is proposing. To calculate stack depth, all you really need is the *maximum* stack depth per instruction. You might slightly over-allocate but it shouldn't really be much of a problem.
Serhiy: what's your use case for all these complicated new features?