Hi Joe, I agree, I think that it should be really "wrong statistics", I don't find any other reason and the results during the night seems to agree with that. I left the code just with the standard configuration during all the night and got different results, a low success rate (<40%) between 1:00 and 4:30, and a highe success rate between 4:30 and 7:00 (>70%), without any change in the code, it might have to do with having more electromagnetic noise in the first period, but I'm not sure about that.
Just adding some more data:
As I have reduced the tx power in one of the nodes and put it two floors above, the reception level is about -85dBm, and I think it is really close to the limit it can receive (lowest level received during the test was -88dB, highest -82dB, average -85dB) so the only way to test is to have a longer run, otherwise any small change in the environment could mask the results. I'll try a longer run test with the different LNA configurations and see what happens, not only a 'quick' ten minutes test.
I changed the power supply of the gateway this morning (it was powered by my laptop during the test at night and replaced it by a mobile charger), and after that I almost get no messages, so the sensitivity to the power noise seems to be present (now I have only 8% success rate). In the receiver I already added the circuit Joe suggested, and after my first test with it I had an high increase on the sensitivity, the lowest level received was about -90dB and went to -102dB. I made it in two different units, and had similar results with both, but now I'm using one of that units I'm not able to get the same low level messages. All the radio initialization is the same (standard initialization, no specific write to RFM69 registers), so I'm not finding a reason for that difference.
At the same time similar units were installed in a vineyard (almost no electromagnetic noise, no houses, factories, nothing close to it), where I'm finding also that the lowest messages received are at about -85dB. All the nodes with received signal level above -80dB are ok, and the success rate starts to decrease fast in nodes that have an average RSSI below -80dB. The lowest level observed at the central node was -88dB, but units with average RSSI below -85dB have a very low success rate (<10%). So this seems to point that the Rx limit is about -85dB (with standard configuration), and that it might be better to work on the central node / gateway (ESP8266 + RFM69HW) to try to improve the reception threshold and/or to reduce the Rx bandwidth in order to improve the reception sensitivity.
When I have more results or identified the cause I'll post it here :-)
Best regards
Fernando