Hm, why do you treat .txt files with crlf conversion as the exception at first? It seems more natural to me to use your second variant, where the rijndael files are the exception, from the start, because those are the files which are somehow special, not the Postgres quickmanual et al...
The code uses a histogram to determine what the default attributes for an extension should be. When the rijndael_*.txt files were added, there were more *.txt files with CRLF than with LF, thus the default to CRLF. When the autodoc documentation files were renamed to *.txt the balance switched.
No, the files should not be crlf converted after the change either. Before the change the EOL had to be CRLF. After the change the EOL had to be LF. Neither variant was EOL agnostic.
As far as I can see, the original test supported either, but Nilsson's commit "No \r in the testdata anymore." removed the support for \r\n.