First note that castor-norm is for direct normalization and in addition your data should be in the histogram format rather than list-mode.
Answering your question, as far as I understand, you need to feed castor-norm the true image and not the reconstructed one by castor-recon. I think you are referring to the following post on the forum.
If you look at this post, it is stated clearly.
It depends on you, whether you want to correct for the attenuation effect or not. Providing atten factors with the normalization is optional and not mandatory. As you can see in the quoted post, if you choose an annular phantom you can probably ignore the effect of attenuation.
To conclude, I think it depends on your specific needs and goal. The more corrections, the better the quality of the reconstructed image.