Consider the output of your original regex and these two variations (I don't know where the  'baz1' comes from in your output):

>perl -wMstrict -le
"my %d = (
     foo => sub { \"[@_]\"; },
     bar => sub { \"(@_)\"; },
 );
 $_ = 'foo fooz bar';
 s/ (\w+) (?(?{ !$d{$1} }) (?!)) / $d{$1}->($1) /xge;
 print;
"
[foo] [foo]z (bar)

>perl -wMstrict -le
"my %d = (
     foo => sub { \"[@_]\"; },
     bar => sub { \"(@_)\"; },
 );
 $_ = 'foo fooz bar';
 s/ (\w+) / $d{$1}->($1) /xge;
 print;
"
Use of uninitialized value in subroutine entry at ...
Can't use string ("") as a subroutine ref while "strict refs" ...

>perl -wMstrict -le
"my %d = (
     foo => sub { \"[@_]\"; },
     bar => sub { \"(@_)\"; },
 );
 $_ = 'foo fooz bar';
 s/ (\w+) / $d{$1} ? $d{$1}->($1) : $1 /xge;
 print;
"
[foo] fooz (bar)
[download]

Then consider the  (?(condition)yes-pattern) construct in the Extended Patterns section of perlre and the fact that, if the null regex is always true, wouldn't the negative look-ahead to the null regex  (?!) always be false? (Coupled with realizing what the logical value of  !$d{$1} is when  $1 doesn't exist in the hash.)