Too many pixels labelled BAD_FIBERTRACE

Looking at the extracted spectra, especially in the n arms, there are far too many pixels labelled as BAD_FIBERTRACE.

Here's a small portion of an n3 quartz, with three fibres labelled:
The reddish pixels are BAD and the cyan ones IPC (which should be fine). And here are the spectra from those three fibres, with the masked pixels shown faintly:
There are clearly many more bad pixels, maybe due to the way that the extractions are done using a matrix inversion?

These were extracted with minFracMask=0.95, so I don't think that there should be any dropouts, there must be something else going on.