Thanks for all the comments (both on the list and direct email).
Many of you suggested the use of edge trigered instead of latches that really works fine in all situations and uses less resources. So, from a design point of view... problem solved. Thanks.
Now, for the inquisitive minds. Just to understand what was happening I was looking in detail into the reports to see what was happening. In fact the equation of a port output bit changed dramatically with and without test point.
As I had a recent design where I use latches without any problem I started to compare the sources to see what was different and the only fundamental difference found is the amount of signals used. On the my previous design (working fine) I used A8 A9 and nWR signals to control the latch while on this one I was using the full A8 to A15 byte plus nWR. So I cut the present design to use just A11 and nWR and bingo it worked always fine, then I added A8 and still works fine, then I added A9... still working fine. Then I added A10 and PUFF!!! stop working.
I looked into the fitter's report and the equations change dramatically, producing something that is clearly NOT A LATCH. I may conclude that I can only have 4 signals controlling the latch.
That is why when I used the testpoint all address lines and nWR were combined to form my test signal and the latch equation generated used this signal plus the A8,9,10 bits so it had only 4 inputs. Without the test point each cell was receiving 9 signals and the synthesised equation was quite different and therefore not working.
If this is a specific limitation of this particular CPLD or software tool an error report or a warning would be very welcome. The reason why only up to 4 signals produce the expected behavior, I think only Altera can tell... (would they bother to comment to this poor web licence user ?!)
I think I'll be another one saying to avoid latches on CPLD designs hi:)
Many thanks again for your help.
Luis Cupido.