Consider the following test case:
#include <iostream>
#include <iomanip>

int main()
{
  double pi = 3.14159;
  std::cout << std::setprecision(40) << pi << std::endl;
}

Compiling and running this program gives the output:
3.1415899999999999  We've got 16 digits after the 
decimal point, even though we've requested 40.

There's code in locale_facets.tcc that limits the precision
if it's "out of range".  I don't see any justification for
doing that either in clause 22.2.2.2.2 of the C++ 
standard, which defines the behavior in terms of printf,
or in clause 7.19.6.1 of the C standard, which describes
printf itself.  (FWIW the C standard does say that there is 
an implementation defined upper limit for printf 
conversions, but it is required to be at least 4095.)