The documentation states that the @ operator suppresses all errors.

However, lines like:
  @ error_reporting E_ALL);
still generate parse errors.

I think that the current behavior is the right thing to do - AFAIK, there is no way to induce a parse error at runtime.  Parse errors in the code that makes up the arguments for eval() and create_function() are caught and handled differently then normal parse errors (at least AFAICT they are - @ suppresses the errors generated in this fashion).

The only real case that I can think of where you might have parse errors at runtime would be situations where you are dynamically including code from untested files -- if you are doing that though, then catching parse errors is probably the least of your worries. :)