This is a feat technically, but will act as more of a fix. The original MATLAB stretched-nmf included a toggle called fillzero that I did not understand or think was important, and therefore didn't implement. Detailed investigation of why we still had worse performance showed this was one of the big reasons.

The fillzero feature is enabled during normalization and means that when components is being stretched, the optimizer treats the area beyond the end of the signal as equaling zero, instead of repeating the final value. When it repeated values, this artificially inflated component intensity at the left and right edges. This, in turn, had cascading effects across the weights and stretch that made them wrong, too. The effect was most extreme during normalization because of the dramatic change in signal magnitude that occurs.

Note that although it looks like one, this isn't intended to be something the user will need to change. It's implemented this way because it matches the MATLAB and fixed normalization issues. In a future PR, I will likely refactor this to be natively part of the algorithm. That will happen as part of a broader cleanup of the normalization process, which currently is entangled in a way that I don't like.