My point was that it requires a pretty expensive key extraction for the ST to be beneficial.

For example, with fileno the ST never gets a look in and it requires a 1000+ files before the GRT starts to make inroads:

c:\test>junk92 *.txt
Files: 46
        Rate   ST GRT1 GRT2 sort
ST    9206/s   -- -21% -27% -58%
GRT1 11725/s  27%   --  -8% -46%
GRT2 12681/s  38%   8%   -- -42%
sort 21855/s 137%  86%  72%   --

c:\test>junk92 *.pl
Files: 657
      Rate   ST GRT1 GRT2 sort
ST   411/s   -- -39% -44% -46%
GRT1 675/s  64%   --  -8% -11%
GRT2 735/s  79%   9%   --  -3%
sort 757/s  84%  12%   3%   --

c:\test>junk92 *
Files: 1293
      Rate   ST GRT1 sort GRT2
ST   186/s   -- -41% -45% -46%
GRT1 315/s  70%   --  -7%  -8%
sort 338/s  82%   7%   --  -1%
GRT2 341/s  84%   8%   1%   --
[download]

So, by the time you get to the point where the standard sort, even with numeric keys, can be beaten, you might as well move to straight to a GRT and have done with it.

    sort => q[
        my @fhs = sort {
            fileno( $a ) <=> fileno( $b )
        } @fhs;
    ],
[download]

Others per your code.