I do wonder, though, whether the RFM69 module, or perhaps even the SX1231H chip itself, is capable of doing any better under even near ideal circumstances. For example, for all I know maybe thermal noise (or something else) dominates the sx1231H below -121dB. I've presupposed that it should be capable of doing better, but an existence proof would at least establish that we aren't hunting for unicorns.
So, as a "Plan B," I'm trying to imagine how to test for that. Notionally, I'm thinking of a brute force experiment something like the following:
1. Run the RFM69 module on its own independent battery power supply, and with all IO done via optical couplers.
2. Connect the DIO0 line to a load switch that can ring a buzzer if it goes high.
3. Put the RFM69 into listen-mode with a very low RSSI threshhold (maybe -122db to start with, as so far -121db is the lowest that I've seen) and a long enough sleep period to provide ample time to execute step #4
4. Initiate listen-mode, and while the RFM69 sleeps, disconnect the optical couplers from the atmega328p, and stick the radio assembly inside a sealed metal can.
5. Time how long it takes for the buzzer to ring (if it rings at all).
If the buzzer rings right away, during the first Rx period of the listen-mode cycle, then I'm inclined to think it would prove nothing more can be done. If it doesn't ring at all, after some TBD number of listen-mode cycles, then I could ratchet down the RSSI threshhold and repeat the experiment until it does, and that would establish the "best case" target to aim for in a more practical design.
It would only be worth doing if it the results were conclusive. Would the above decide the matter, or would there remain room for doubt?