Two points from an earlier post:
- you don't need to include 28F512_CS into the ports' decoder (hence improving timing budget); if you really need to write into 28F512, you can do it using a dedicated pin (e.g. from port 1 or 3) manually - the write is already slow, this won't worsen it (at the same moment, you don't need to decode addresses for 28F512 at all)
- using 74xx259 you have 8 decoded per chip, no extra glue logic, no extra delay

And I would not draw a bare CPLD output through a "flat cable, max. 2m long".

Jan Waclawek