His string eval was missing calls to quotemeta. I fixed this, added a few more solutions, and reran the benchmarks. They show that:

• Without using eval (rot13dynre), your task can be accomplished in 2/3rd of the time it currently takes (rot13splitmap).

  my %rot;
  @rot{'A'..'M'} = ('N'..'Z');
  @rot{'N'..'Z'} = ('A'..'M');
  @rot{'a'..'m'} = ('n'..'z');
  @rot{'n'..'z'} = ('a'..'m');
  
  my $chars_to_change = join('', map quotemeta, keys %rot);
  
  sub rot13dynre {
     local $_ = @_ ? $_[0] : $_;
     s/([$chars_to_change])/$rot{$1}/g;
     return $_;
  }
  [download]

• By using eval (rot13trbuilt), your task can be accomplished in 1/167th of the time it currently takes (rot13splitmap).

  my %rot;
  @rot{'A'..'M'} = ('N'..'Z');
  @rot{'N'..'Z'} = ('A'..'M');
  @rot{'a'..'m'} = ('n'..'z');
  @rot{'n'..'z'} = ('a'..'m');
  
  *rot13builttr = eval 'sub {
     local $_ = @_ ? $_[0] : $_;
     tr{' . join('', map quotemeta, keys   %rot) . '}
       {' . join('', map quotemeta, values %rot) . '};
     return $_;
  }';
  [download]

That should help convince your peers that this limited, well controlled, easily testable use of eval is appropriate here.

Benchmark results:

Benchmark:

Update: Reorganization for clarity.