It was a mix of problems, I'll post here hoping it might be useful to others. My first observation with the sniffer is that each node transmits always the three retries, and only the last one was acknowledged, this is why the sketch 'sort of' worked, using all the retries. The problem was introduced by some code I added during the previous debug phase (when i had communication problems with the RFM69 module), logging the messages to a server before sending the acknowledge. Now I make a copy of the received message, send the ack and only then process the message and I'm observing the ack delay at about 10ms. As Joe said, there is some 'jitter' in this reply time probably due to the other activities (WiFi and TCP stack) running on the ESP, in a quick tet I have observed from 7ms up to 14ms, something that with the ATMEGA328 as gateway was much more deterministic, but nevertheless is ok, since the default timeout is 40ms. This reply time only worried me because I have to keep the node awake until it receives the acknowledge, so more than 100ms was bad, now with about 10ms it is ok! Thanks again for your help!
Best regards
Fernando