Provisionally accepting your distinction between atheism and agnosticism, in what way is the former useful and the latter not?

The results of all these tests point out that falls are not lethal, of course :-P

That's where an untested auxiliary belief figures in - "if something hurts in proportion to variable x (i.e. the height of the drop), experiencing that thing when x is very large will probably kill you".

That's basically the Duhem-Quine spiel right? Which is why strict falsificationism doesn't quite work. But that's not to say a weaker form of falsificationism can't work: a network of ideas is useful to the degree that nodes in the network are testable. A fully isolated network (such as a system of theology) is useless.