Here are my findings:
1) No CosmeticCorrection
When the calibrated light frames are debayered and then registered with default parameters, the alignment is on the hot pixels. This can be fixed by setting 'Hot pixel removal' from 1 (the default) to 2 in StarAlignment. In ImageIntegration I tried several rejection algorithms in order to remove the hot pixels. In the order of decreasing effectiveness:
Winsorized Sigma Clipping (Sigma_low=4.0, Sigma_high=2.5; slightly more effective than Sigma Clipping, but the fraction of rejected values was higher as well)
Sigma Clipping (Sigma_low=4.0, Sigma_high=2.5)
ESD (default parameters)
Linear Fit (lf_low=5.0, lf_high=4.0)
So winsorized Sigma or Sigma Clipping seemed suited best. However, even with the shown settings, slight remnants of hot pixels remained detectable in the integration despite a fraction of rejected values in the range 1 - 1.5 %.
2) CosmeticCorrection applied on CFA data
As stated in this thread, applying CosmeticCorrection (Auto detect, Hot sigma) on the calibrated light frames in CFA format (option CFA enabled) yielded strange results. I gradually lowered Hot Sigma until there was erosion of the stars detectable. At a Hot sigma of 1.6 this was just not yet the case. This rendered possible that StarAlignment worked OK with 'Hot pixel removal' set to the default value of 1. However, applying rejection algorithm ESD with default parameters showed walking noise clearly.
3) SplitCFA, then CosmeticCorrection applied on the monochrome images
The workflow SplitCFA, application of CosmeticCorrection to the monochrome images, register, integration to R, G and B monochrome images, drizzle integration of the separate channels and combination of the R,G and B channels with ChannelCombination led to the best results. Using rejection algorithm ESD with default parameters resulted in a very small fraction of rejected values (around 0.10 - 0.12 % for both rejected low and rejected high values), and the hot pixels were removed completely.
So my conclusion is that the appropriate application of CosmeticCorrection is the most efficient way (in terms of preserving the highest SNR) to avoid the streaks in the integration. Unfortunately, currently applying CosmeticCorrection on the calibrated light frames in CFA format did not work satisfactory with these data. The approach using of splitting the CFA data to monochrome images worked fine, but is rather tedious. With these data, I was not able to reproduce your finding that omitting CosmeticCorrection leads to an integration devoid of walking noise. To my opinion this is only possible when pixel rejection is applied rather aggressively, and that will lower the SNR. The linear fit algorithm with parameters lfit_low=4.0, lfit_high=2.0 (your settings in post #36) lead to the following total fractions (low + high) of rejected values: R: 5.5 %, G: 5.7 %, B: 5.3 %.
Bernd