Just to make sure I am clear on this, you want to know which is better in terms of making maps?
1) First take an ungeoreferenced raste image, georeference it and then make a map; or 2) Make the map and then georeference that map?
I think it would make sense to first georeference the data (raster image) and then make the map from that.
If you will be adding information to the map (digitising or adding in existing data) this data will be correctly positioned spatially. This means that in the future, if you need to re-use this information it is correctly positioned spatially. If you make a map and add information to it then: a) You digitised data would be incorrectly positioned spatially and trying to put it in the correct place can be time consuming and inaccurate b) Any existing information that you have will not line up / match the map correctly.
I hope I have understood and answered your question.
Definitely georeference your images first. Then capture the vectors from the rectified image(s)
In doing several data capture projects I found this reduced the work significantly.
1. You do not have to reproject or adjust the vector data 2. There are no edge matching issues because you can just add more images to cross edges, or merge multiple images into a single large image. This was very useful for utilities where you are capturing manholes and joining them with pipes across sheet boundaries with a single feature. 3. All the issues of datums in the image and other distortions can be handled simple with the georeferencing links. 4. For QA work just supply the rectified images that can be displayed under the final work.
In one case the controls for registering the images were marks on manholes that had been surveyed in the field. When these were added a check could be made with the RMS errors to ensure that they were consistent and the registration was correct.
[In the old days when we captured data using digitiser tables the opposite was true. We had to capture the vectors in "page units" and then transform to "map units" using the corner TICS. Then a painful edge matching process for each sheet and separate capture of the attributes. Finally a projection to the desired projection if different from the source.]