Update: While the  (?: (b.) (?{ push @two, $2 }))+ approach discussed in this reply can certainly work, I think I would prefer the two-step approach outlined in choroba's reply above for its greater readability and maintainability.

If I take the for-loop out of the equation, I get something closer to what I expect (still Strawberry 5.14.4.1):

c:\@Work\Perl\monks>perl -wMstrict -le
"my $string = 'ab1b2b3b4d';
 my @two;
 my @matches = $string =~ /(a)
                           (b.(?{ push @two, $2 if defined $2 }))+
                           (d)/x;
 print qq{'$string' -> all:(@matches) -> 2:(@two)};
 ;;
 $string = 'ab5b6b7d';
 @two = ();
 @matches = $string =~ /(a)
                        (b.(?{ push @two, $2 if defined $2 }))+
                        (d)/x;
 print qq{'$string' -> all:(@matches) -> 2:(@two)};
 ;;
 $string = 'ab8b9d';
 @two = ();
 @matches = $string =~ /(a)
                        (b.(?{ push @two, $2 if defined $2 }))+
                        (d)/x;
 print qq{'$string' -> all:(@matches) -> 2:(@two)};
 ;;
 $string = 'abxd';
 @two = ();
 @matches = $string =~ /(a)
                        (b.(?{ push @two, $2 if defined $2 }))+
                        (d)/x;
 print qq{'$string' -> all:(@matches) -> 2:(@two)};
 ;;
 $string = 'ad';
 @two = ();
 @matches = $string =~ /(a)
                        (b.(?{ push @two, $2 if defined $2 }))+
                        (d)/x;
 print qq{'$string' -> all:(@matches) -> 2:(@two)};
"
'ab1b2b3b4d' -> all:(a b4 d) -> 2:(b1 b2 b3)
'ab5b6b7d' -> all:(a b7 d) -> 2:(b5 b6)
'ab8b9d' -> all:(a b9 d) -> 2:(b8)
'abxd' -> all:(a bx d) -> 2:()
'ad' -> all:() -> 2:()
[download]

And if I re-arrange the capturing, I get exactly what I expect:

c:\@Work\Perl\monks>perl -wMstrict -le
"my $string = 'ab1b2b3b4d';
 my @two;
 my @matches = $string =~ /(a)
                           (?: (b.) (?{ push @two, $2 if defined $2 })
+)+
                           (d)/x;
 print qq{'$string' -> all:(@matches) -> 2:(@two)};
 ;;
 $string = 'ab5b6b7d';
 @two = ();
 @matches = $string =~ /(a)
                        (?: (b.) (?{ push @two, $2 if defined $2 }))+
                        (d)/x;
 print qq{'$string' -> all:(@matches) -> 2:(@two)};
 ;;
 $string = 'ab8b9d';
 @two = ();
 @matches = $string =~ /(a)
                        (?: (b.) (?{ push @two, $2 if defined $2 }))+
                        (d)/x;
 print qq{'$string' -> all:(@matches) -> 2:(@two)};
 ;;
 $string = 'abxd';
 @two = ();
 @matches = $string =~ /(a)
                        (?: (b.) (?{ push @two, $2 if defined $2 }))+
                        (d)/x;
 print qq{'$string' -> all:(@matches) -> 2:(@two)};
 ;;
 $string = 'ad';
 @two = ();
 @matches = $string =~ /(a)
                        (?: (b.) (?{ push @two, $2 if defined $2 }))+
                        (d)/x;
 print qq{'$string' -> all:(@matches) -> 2:(@two)};
"
'ab1b2b3b4d' -> all:(a b4 d) -> 2:(b1 b2 b3 b4)
'ab5b6b7d' -> all:(a b7 d) -> 2:(b5 b6 b7)
'ab8b9d' -> all:(a b9 d) -> 2:(b8 b9)
'abxd' -> all:(a bx d) -> 2:(bx)
'ad' -> all:() -> 2:()
[download]

Update: The expression
    (?: (b.) (?{ push @two, $2 if defined $2 }))+
works just as well without the statement modifier:
    (?: (b.) (?{ push @two, $2 }))+
I don't understand the use of the statement modifier: Could  $2 ever be undefined in this situation?