Your benchmark is including the loop overhead - which is appropriate for "real" code, but dilutes the difference between the two access methods. Altering the appropriate parts of your code to:

my $lookup = '
    _____
';
$lookup =~ s/_____/'$a = $contracts{1}[0];' x NUM_ACCESSES/e;

my $refit = '
    my $r = $contracts{1};
    _____
';
$refit =~ s/_____/'$a = $r->[0];' x NUM_ACCESSES/e;
[download]

and using your last test set I get:

           Rate lookup  refit
lookup 245666/s     --   -32%
refit  360597/s    47%     --
[download]

which is double the difference shown with the loop overhead included, and is comparable with the result from my benchmark.

Ain't benchmarks fun, almost as good as statictics! :-D