Some people misrepresent the halting problem and its corollaries as "we can under no circumstance predict any aspect of the behaviour of any program", which is in fact wrong.
What is true is that, for every property we may want to assert about a program, there exists a program (on a Turing machine or in a Turing-complete language, so excluding any of the total programming languages) for which you will fail to tell if it has that property.
Not really, as a turing machine is fully able to simulate any other turing machine.
The problem is not that the "termination checker" may not run the program it's checking (it is allowed to do that). The problem is that if that program runs forever then the "termination checker" program will also run forever and thus never return "yes" or "no".
(You could say, well just output "no" if the program runs for longer than x days / weeks / months / years, but it's quite easy to write programs that run for a very long time but do eventually terminate. At best, you'll manage to write a checker that can output "yes" or "I don't know").