I had a look at it. The formatting seems correct, also the curves are displayed in the spectral unmixing module.
However, many of them are quite similar. I would suggest that you only use one of "Veg", Bright" and "Dark" in the end. It should not alter the quality of the output as all three classes show very different spectral profiles. But this probably will not solve the problem.
I also found some NAN values in the table, maybe they cause a distortion:
My suggestion would be to skip Dark_1-3 and only use Dark_4
Lastly, you could try to change the model to "unconstrained" and see if anything changes. Classes won't sum up to 1 any longer but maybe you get results you can work with.