I tried to reproduce the described behaviour.
What I’ve observed is that the NaN values are only introduced when trying to reproject to the same CRS, and doing this does actually make no sense.
I’ve tried with a resampled S2 MSI L1C product which is resampled to 60 meters. The CRS is already UTM 33N.
If reprojecting to Geographic Lat/Lon (WGS84) everything looks good. Reprojecting to ETRS89 / UTM zone 33N works, also reprojecting to UTM 32NN works. What does not work is to reproject to UTM 33N, and this is the original CRS. While this could be considered a bug it is not really serious because if you want to use this CRS you don’t need to do anything.
More serious is, that the classification seems to require the input data to be provided in geographic lat/lon. I’ve now created an issue for this.