Perhaps you missed the significance of the timings:

C:\test>1011850 -O=100000
Dict read at C:\test\1011850.pl line 96.
objects loaded at C:\test\1011850.pl line 106.
App created at C:\test\1011850.pl line 112.

Ovid: (optimal ordering) returned 9193 objs in 4.447958 seconds
Ovid: (pessimal ordering) returned 9193 objs in 7.236690 seconds

RichardK: (optimal ordering) returned 9193 objs in 4.065038 seconds
RichardK: (pessimal ordering) returned 9193 objs in 6.219226 seconds

BUK: returned 9193 objs in 0.760572 seconds
Filters run at C:\test\1011850.pl line 138.
[download]

Running your OP algorithm against a set of 100 filters each carefully arrange to filter an increasing number of objects and ordered best to worst, and running two tests -- optimal and pessimal -- the difference between then is less that 50%. That is the best you can hope to achieve with your order-the-filters mechanism.

However, by altering the way the filters operate upon the objects, you can achieve an order of magnitude saving.

Or perhaps you've just decided how you want to tackle the problem. Regardless...