Yes, fitness is related to the environment. But...

In this case, if the algorighm becomes simpler if FITNESS is more tightly associated with the INDIVIDUAL. For example, this:

  sub choose {
    my $self = shift;
    my $f = rand(1.0);
    my $index = 0;
    my $sum = 0.0;
    foreach my $fitnes (@{$self->{FITNESSES}}) {
      $sum += $fitnes; 
      return ${$self->{INDIVIDUALS}}[$index] if $sum >= $f;
      ++$index;
    }
    die "can't select an individual";
  }
[download]

becomes:

  sub choose {
    my $self = shift;
    my $f = rand(1.0);
    my $sum = 0.0;

    foreach my $individual (@{$self->{INDIVIDUALS}}) {
      $sum += $individual->{FITNESS}; 
      return $individual if $sum >= $f;
    }

    die "can't select an individual";
  }
[download]

which I consider an improvement, not only due to the fact that there's (a little) less code, but the concept of choosing an individual and not a fitness is emphasized.

Finally, I must admit that I have a bias in relation to the idea you presented. Yes, environment does determine fitness. However, what if you're trying to evolved generalized behavior, i.e., a program that will perform well in any environment, and not simply the one's it was trained in? The little bit of work I've done with GP has been focused in the direction of trying to avoid such "over-training" or specialization.