Thanks for your comments.

yes it is

Do you mean that in rare cases the internal Reset logic can ignore an external reset signal?

Which are the 10% of the recommendations you have not followed?

-I don't enclose the board or CPU in a metal enclosure in order to protect it from EMI traveling in the air.
-I don't have GND fillouts on top solder.
-My board doesn't have a ferrite at power inputs.

I do believe that you have an additional problem when the watchdog does not bite, I am fortunately not familar with the Atmel chips, but if there is ANY method to disable the watchdog or the feeding is anything else that two specific consequtive writes the chance that runaway code will feed it or disable it is there.

Well, I am using an external watchdog since I don't trust much the internal WD. There is no way for the CPU to externally disable the watchdog. In few cases I had chance to analyze the circuit with an oscilloscope and to my surprise while the Reset Pin(9) was being bitten by continuous external WD pulses (duty cycle of 10ms) the CPU was refusing to restart! The only way to restart it was resetting the power source.

I think it is a specific problem with ATMEL where under certain conditions the internal reset logic fails to monitor external reset requests.