I'm not a big fan of the glob trick - there's always the possibility something matches (although for the given pattern, that's pretty unlikely). But just for kicks, I compared your solution with mine:

$ /usr/bin/time ./perlmouse ; /usr/bin/time ./ray_johnson 
-9 + 8 * 7 + 654 * 3 - 2 - 1 == 2006
 9 + 8 * 7 + 654 * 3 - 21 == 2006
 9 + 8 * 7 * 6 * 5 - 4 + 321 == 2006
 9 * 8 - 7 + 654 * 3 - 21 == 2006
56.60user 0.19system 2:04.01elapsed 39%CPU (0avgtext+0avgdata 0maxresi
+dent)k
0inputs+0outputs (356major+87minor)pagefaults 0swaps
-9+8*7+654*3-2-1 == 2006
9+8*7+654*3-21 == 2006
9+8*7*6*5-4+321 == 2006
9*8-7+654*3-21 == 2006
56.98user 3.45system 2:43.21elapsed 41%CPU (0avgtext+0avgdata 0maxresi
+dent)k
0inputs+0outputs (398major+21556minor)pagefaults 0swaps
[download]

The running time is about the same, but your solution uses a lot more memory, as it makes a list of all possible expressions to test before making the first test, while my solution tests each expressions right after it is constructed. My solution has a lot less pagefaults.