That example is not self-modifying, though. The BIT absolute instruction simply uses the LDY immediate instruction as operand data by means of overlapping them.

That example is not self-modifying, though. The BIT absolute instruction simply uses the LDY immediate instruction as operand data by means of overlapping them. This saves some bytes as you'd otherwise have to use relative branching (and there's no branch-always) or a three byte jump to load a different Y value, but the code remains unsabotaged. These tricks are IMO an indicator that they had a good idea that the game code was going to be a tight fit (or had possibly already discovered that the hard way), so they were optimizing for size. A jump would be faster than executing the BIT instruction. The manually loading an address to the stack and running RTS is an interesting trick. Wozniak used it in his SWEET-16 interpreter, pushing the interpreter loop entry point to the stack before jumping to the subroutines that execute the byte code, just so that he could use RTS to return to the interpreter loop. In that sense it fulfills an entirely different purpose than a JMP indirect and also saves some bytes, again at the expense of execution speed and non-gray hair. For a table based interpreter, it fills gap left by the lack of a JSR indirect instruction.