Linux – How is the open file table structured

The behaviour you're looking for is a special case:

cp -R [-H|-L|-P] [-fip] source_file... target

[This] form is denoted by two or more operands where the -R option is specified. The cp utility shall copy each file in the file hierarchy rooted in each source_file to a destination path named as follows:

• If target exists and names an existing directory, the name of the corresponding destination path for each file in the file hierarchy shall be the concatenation of target, a single <slash> character if target did not end in a <slash>, and the pathname of the file relative to the directory containing source_file.
• If target does not exist and two operands are specified, the name of the corresponding destination path for source_file shall be target; the name of the corresponding destination path for all other files in the file hierarchy shall be the concatenation of target, a <slash> character, and the pathname of the file relative to source_file.

It shall be an error if target does not exist and more than two operands are specified ...

Therefore I'd say it's not possible to make cp do what you want.

Since your expected behaviour is "cp -r dir1 dir2 (when dir2 already exists) would remove the existing dir2 (and any contents) and replace it with dir1":

rm -rf dir2 && cp -r dir1 dir2

You don't even need to check if dir2 exists.

The rsync solution would be adding a trailing / to the source so that it doesn't copy dir1 itself into dir2 but copies the content of dir1 to dir2 (it will still keep existing files in dir2):

$ tree dir*
dir1
└── test.txt
dir2
└── test2.txt

0 directories, 2 file
$ rsync -a dir1/ dir2
$ tree dir*           
dir1
└── test.txt
dir2
└── test.txt
└── test2.txt

0 directories, 3 files
$ rm -r dir2          
$ rsync -a dir1/ dir2
$ tree dir*           
dir1
└── test.txt
dir2
└── test.txt

0 directories, 2 files

"history of sync" is a tough thing to google for ;-) A totally uninformed guess is that at least one legitimate use is for kernel developers making changes to the sync(2) system call to test. They need a way to invoke the syscall that's a little more convenient than halting or umounting a filesystem?