Beman Dawes
After doing a lot of reading of git docs and about how others dealt with similar problems, that was also my conclusion. I'm wondering why others haven't suggested renormalization?
They have.
Since we can't fix these files in historical versions we could re-run the conversion with gitattributes that unsets the text attribute for the problematic files (something like "*.bat -text", "*.vsproj -text", "*.svg -text"). Then in git update the gitattributes (to something like "*.bat text eol=crlf", "*.vsproj text eol=crlf", "*.svg text") and normalize the files so that it does what we want. We can script that, so it shouldn't require too much work.
As long as we are sure the .gitattributes are correct, and then renormalize, why do we have to rerun the conversion at all? What am I missing?
It depends on whether you want to "correct" history or not. If you want to go back in time and renormalize, then all the SHAs in submodules change, and the branches in the master repo need to be rewritten accordingly.