I am pretty sure (though have not vetted) that triangulation of LoRa is possible even at very low SNR.

The trick is understanding LoRa's trick, which is simply to "skew" the signal across time (via chirps), modulo a window of the configured bandwidth around the center frequency. The key is that the skew rate is purely a function of the spreading factor, bandwidth, and IQ polarity (= pol × BW² / 2^SF), so there's a small-ish finite number of skew rates. So you can just modulate raw IQ data with carriers at each of these skew rates to find one which gives you a bunch of carrier waves that hop around discretely at about twice the symbol rate, looking like an FSK signal. You can then bin this at a factor of, say 2^(SF-2) to correlate the signal and raise it up above the noise floor, on which you can apply any standard triangulation technique.

I'll try vetting this soon and reply to this post with results.

Sounds like you know what you’re talking about. I’d be very interested in learning more. I’m not a radio or even software expert. I’m a data scientist with a math background. Interested in communicating across mesh networks anonymously.

I know only basics about mesh networks per se. I'm speaking here from a background in signal theory / amateur radio / research on the LoRa protocol. If you have specific questions on those aspects feel free to contact me, my e-mail is in my profile.

Is your goal nondetection? If so, know that triangulating a radio signal is fairly straightforward, even in a dragnet fashion. The exception I think is something like cryptographic DSSS (found in military use) wherein not only is the signal far below the noise floor, but the spreading function is not predictable by an adversary (LoRa being very much predictable). Even then, you're limited by physics to only be able to transmit so far without the signal being detectable near the transmitter.