Doesn't this change the behavior in optimized code, for assert statements with side effects?
That's why I said it will be a long discussion :-) I would also propose the elimination of the -O option, and the notion of optimized code. People who want to save the time for execution of the assert statements could use some kind of source transformation, which might be easy to implement with lib2to3 (or a generalization thereof).
If we don't have specialized syntax -- that is, if 'assert' just becomes another built-in function -- then it can't be special-cased in the translation to byte-code, can it?
It still could. First, assert may continue to be a statement, even though it is invoked as a function call, similar to None, True, False. In that case, the compiler could still eliminate it, or (IMO preferred), wrap it into a conditional statement that checks whether assertions are disabled, and, if so, jumps over the assertion. Even if it isn't a keyword, the compiler could still special-case it, and add a run-time check to determine whether assert is still bound to the built-in assert. Regards, Martin