The change that I thought was the most-likely cause of the problem is
the confirmed culprit, even though when I tried to revert to the
lehuakona version of the offending routine from the development system,
I couldn't generate the bug. Going the other way (starting from the
lehuakona source) I tried substituting the various MFITTREND changes and
only after switching the culprit routine did the fits look sensible
across the whole cube.
To confirm: it only comes into play for method=global in lehuakona.
The affect may vary with the NDF in particular the dimension of the
spectral axis. So I suggest you home in on the NGLS processing that
used this option.
It was a revised API for one routine not being applied in all of the
many places where the routine is invoked. This oversight was not found
until five weeks post lehuakona's release, while tracking down another
bug for Brad. I discovered the API error from just looking at the code
rather than from the output NDFs of MFITTREND.
Brad, do you cherry pick fixes for patches, or do you expect each of us
to make a case that a given bug fix is important? Is there a time
limit?
Malcolm
|