You can think of the overhead of recursion as either useless overhead or useful debugging information. As soon as someone implements tail recursion optimization, the output of caller becomes massively more confusing. If you further add the complexity of continuations, good luck producing something resembling a stack backtrace which is both accurate and comprehensible.

Personally I tend not to think of the overhead of recursion. If the algorithm is more clearly coded for me recursively, I just do it and worry about optimization later if I need to. (Later usually doesn't come.)