I think it is working correctly. Let’s consider a 1K time series:
1K real samples from ADC.
Converted to 2K complex samples by adding a zero to every other sample. Now we have 2048 samples
FFT on 2048 complex samples. This gives a mirrored spectrum. Take the lower half gives 1024 complex, and take the magnitude of those give 512 reals.
So, a 1K time series of reals sampled at 48K ultimately produces 512 bins between 0 and Nyquist, but 1024 bins overall. The resolution between 0 and Nyquist is Fs/N, or 48K/1024 = 46.88 Hz. And 46.88 Hz * 512 = Nyquist
What software version are you running? Has that changed? I don’t think there’s been a change to export. But I think the current behavior you are reporting is expected.
A source of confusion perhaps is the “FFT” reading on the display. That can be ambiguous as it’s not clear if taking about input complex, input reals, output complex output reals, output half spectrum, etc.