Thanks both kjensen and ksa for your answers. Gateway is running in legacy mode, ie, it forwards all xbee network traffic to its serial port that is forwarded to a tcp socket, and all the traffic sent to the tcp socket is forwarded to the serial port and then to the xbee network.
On the other tcp socket side there are a set of scripts that manage the data coming and going to the network. Very similar to the python running in the coordinator model, but in my case running in other host.
And I'm pretty sure that the fact that being in legacy mode is related to the error. From six devices in sleep mode, there's one that always falls from the network, it can last one day or one month, but finally it becomes down. If, in this situation, I reboot the coordinator nothing happens, ie, the end-device keeps on out of the network. But if I reboot the device in non-legacy mode, coordinator sees the end-device as soon it send the first read. Once added to the end-devices list I reboot the coordinator in legacy mode again and it is running well from days.
I post this matter some time ago in this forum, but I didn't get any help. I'm quite sure that it must be a bug, ie, something that the coordinator performs in normal mode is not performed in legacy mode.
It's important to state that the other five end devices work perfect and that all share the same firmware and configuration. It's important to state too that this end-device is located at the side that other end-device working fantastic.
Finally I changed the end-device. I was pretty sure that the cause was not the whole device, but anyway...
I've read many times the manual (XBee - xBeePro ZB RF Modules), some books (Building Wireless Sensor Networks) and this forum, but it seems that the configuration is right. It's true that the attached configuration was a late modification where I increased the node timeout (SN: 90 in coordinator/routers intead of SN:30), but before this change the SN was 30 in coordinator/routers and the problem remained.
Very boring story, I have to admit. But any help or experience in similar environments would be really apprecianted.