my $packed = join '', unpack '(a3x)*', pack 'l*', @sint24s;
[download]

Rather than the above,
    my $packed = pack '(l<X)*', @array;
is more concise, probably faster (but this latter not checked).

A shame there seems to be no way to get rid of the final  map on the unpacking side.

>perl -wMstrict -le
"use Test::More 'no_plan';
 use List::Util qw(shuffle);
 ;;
 sub rand_s24 { return map { int(rand 2**24) - 2**23 } 1 .. $_[0]; }
 ;;
 use constant MIN_S24  => -(2**23);
 use constant MAX_S24  => -MIN_S24() - 1;
 use constant VALUES   => (0, 1, -1, MIN_S24, MAX_S24);
 use constant N_VALUES => scalar(@{[ VALUES ]});
 use constant MAX_LEN  => 25;
 ;;
 for my $pass (1 .. 200) {
   for my $len (1 .. MAX_LEN) {
     my @to_pack = (rand_s24($len - N_VALUES), VALUES)[-$len .. -1];
     @to_pack = shuffle @to_pack;
     $len == @to_pack or die qq{bad len: array to pack};
     ;;
     my $packed = pack '(l<X)*', @to_pack;
     length($packed) == 3 * $len or die qq{bad len: packed string};
     ;;
     my @unpacked =
       map { unpack('l<', qq{\x00$_})/256 } unpack '(a3)*', $packed;
     $len == @unpacked or die qq{bad len: unpacked array};
     ;;
     is_deeply \@unpacked, \@to_pack,
               sprintf('pass %d: len %d', $pass, $len);
     }
   }
"
ok 1 - pass 1: len 1
ok 2 - pass 1: len 2
... (4996 lines elided) ...
ok 4999 - pass 200: len 24
ok 5000 - pass 200: len 25
1..5000
[download]