The problem is that such usage implies a degree of “magic” resulting from overlooking what is actually going on.
If I had, say, an AM receiver with 6 KHz IF filter, there’d be a certain amount of noise in the passband, eg, starting with -174 dBm/Hz, adding 10log6000 yields -136 dBm and add in another 10 dB for the receiver imperfections known as “noise figure” - so there’s a noise power of -126 dBm.
But now if I decide I’m going to use morse code, I can switch in a 600 Hz filter, and have 10dB less noise in that narrower bandwidth -136 dBm.
If I have distant signal that’s actually delivering -130 dBm to my receiver, is that above the noise or below?
If I looked at it in terms of the AM receiver, it would be below the noise floor. But if I look at it in terms of the the receiver I would actually use, then it’s above the noise floor.
That’s all very traditional as it deals strictly in terms of frequency.
But the situation isn’t ultimately different with LoRa, it is just that the differentiation of signal from noise is not made on the basis of a static bandwidth, but rather on recognizing what looks like a LoRa chirp.
Evaluate the noise at a LoRa receiver as if it were a type of a receiver it is not - which is to say ignore everything after the IF filter - and sure, it receives signals below the noise floor.
But evaluate it in terms of what it is actually looking at, and no - it only receives signals that are above the noise level of things that look like signal but are not.
Something like Sigfox also receives signals which are below the noise power at the IF filter of a LoRa receiver tuned to the same frequency; but it does that by applying a far narrower IF - eg not 125 KHz, but something well under a KHz - that’s at least a 20 dB reduction in noise, too. Of course it does create the challenging of finding a signal that is narrower than the error of the transmitter frequency source, just as LoRa has the challenge of needing to see things that zig and zag, but not knowing the schedule on which they will start to do so…
The redundancy of 4/5 coding could also, if one cared to do the math, be expressed in dB.
There’s a definitely a cleverness in these systems, in that they manage to change what is seen by the decoding part of the receiver in terms noise to something other than that of a static fixed bandwidth. But they only change the style of the game, not its rules - signal is still signal, not-signal is still noise, and the ratio between them is math.
(In fairness, the “morse code” example is also overlooking something, which is the final filtering is the brain of the skilled operator which can allegedly also manage to receive things at or below the noise floor of what escapes the narrow IF filter, and then go back and apply error correction over missing symbols)