Neither a smaller radius nor fewer swaths nor both together is sufficient to prevent the steps in the extraction.

Have you quantified by how much the neighbouring fibers are contaminated ? Yabe-san figure shows quite high level.
I looked at the only 1-every-4th fiber profiles we have from november :
The first neighbours show contamination less than 1%.

I don't believe this is an issue with contamination: a profile with a radius of two pixels should have zero overlap with neighbouring fibers. Rather, I believe there's a problem with how the profile is used in the extraction.

Well, I agree there is no overlap in the profiles themselves.
But again, if my understanding is correct, so far, as a good approximation, we considered that each pixel is some flux measured from one object only right ?
But because of the wings, each pixel is actually the sum of multiple objects with some weights.
The problem is we don't know those weights, and until we do there is nothing much we can do.
That's what I called contamination, the fact that if you have an bright object that peak at 20k in one fiber, then the pixels from neighbouring fibers will get an extra 200 counts on top of their own object.

Sure, but that can't produce the strong steps in the extracted spectrum.

The steps in the spectra are remedied if the profiles go to zero. Currently, they don't because they were measured on full-fiber-density images and therefore have small radius, so the outermost profile samples are signficantly above zero. If I force them to zero (by subtracting the average of the outermost samples), the steps disappear. The attached plot shows the same fibers as before, but from a pfsArm.

Updated the CALIB in /work/drp at Subaru and in /projects/HSC/PFS/Subaru on Tiger at Princeton.

Paul did some experiments, but without actual improvement.
Again, profiles are now generated differently (PIPE2D-907), should be wider and less subject to that kind of issues.