Yes, if you have pushed the return address, perhaps as a tail-call optimisation (i.e. call to a function which links to a leaf function), then you can jump through the gate to save the size of the return instruction.

And if you simply execute a JMP instead of a required CALL from a main(empty stack) routine? That will go wrong on RETF?

Also, JMP doesn't change CPL, while CALL does?

A jmp though a call gate cannot change privilege. Basically it's a useless functionality.

Just improved privilege levels for MAX(CPL,RPL)<=DPL requirement to apply to non-code segments as well(as wrll as removing the JMP specific check(might need reimplementing it on call gates). Might even remove that specific part altogether because it's already handled further upwards.