Preserve attributes in syncfiles without disturbing parent directories

[Obihoernchen]

Problem

The syncfiles rsync push carried no preservation flags:

rsync -rvLD <staging>/ root@[node]:/

Without -p, rsync explicitly disables setuid/setgid/sticky on new files, so a syncfile entry could never deliver
a setuid binary, and existing files on the node kept whatever mode they already had.

Why the flags could not just be added back

It was tried once in e52a9ff7
("Have syncfiles attempt to preserve more") and rolled back in c0287e93:
"Unfortunately, it will try to change parent directories on the way to the files actually being written."

The staging tree mirrors the full destination path, so -r walks it as real directory entries and rsync applied
the staging directory's attributes to the real /etc on the node. Naming every staged file through --files-from
plus --no-implied-dirs confines preservation to the content actually being synchronized:

rsync -vLpgotDA --no-implied-dirs --files-from=- \
      --usermap=<confluent uid>:root --groupmap=<confluent gid>:root \
      <staging>/ root@[node]:/

What changed

• The file list is built in memory and fed to --files-from=- on stdin. Otherwise empty directories are named
  explicitly, since nothing else would pull them across.
• util.check_output gained a keyword only input=, mirroring subprocess.run. stdin is only piped when
  input is given, so existing callers are unaffected.
• mkpathorlink copies the source directory's attributes onto the staged copy. Files are staged as symlinks and
  rsync reads through to the real file, but directories are staged as directories.
• The confluent account's ownership is mapped to root via --usermap/--groupmap, since it normally does not
  exist on the node. Every other owner still matches by name, so this is not a blanket --chown.
• --xattrs from the earlier attempt is left out: with --copy-links rsync reads xattrs off the symlink, not its
  referent, so it transfers nothing here while adding a failure mode on hosts without xattr support.

Warning

This changes existing syncfiles behaviour. Attributes that used to be left alone are now taken from the
source on every sync:

• Permissions are asserted every time. Existing files on the node previously kept their own mode, so a local
  chmod on a synced file survived. It is now reverted on the next sync.
• Ownership comes from the server. Synced files used to land root:root regardless of the source. A file
  owned by a named account now arrives owned by that account, matched by name, falling back to the server's
  numeric id when that name does not exist on the node. Only the confluent account is mapped to root.
• Timestamps and ACLs are carried over (-t, -A), where previously they were not.

Directories that are only traversed on the way to a synced file remain untouched, which is what the second test
table below covers.

Testing

End to end against a deployed AlmaLinux 10 node, same fixture before and after.

Files and directories delivered by the sync:

synced item	path	source on server	before	after
setuid binary	/opt/sync/sub/setuidapp	4755 root:root	755	4755 root:root
setgid binary	/opt/sync/sub/setgidapp	2755 root:root	755	2755 root:root
existing file	/opt/sync/sub/tightmode	0640 root:confluent	666, stale mode kept	640 root:root
confluent owned file	/opt/sync/sub/cflowned	0640 confluent:confluent	640 root:root	640 root:root
system user file	/opt/sync/sub/daemonowned	0644 daemon:daemon	644 root:root	644 daemon:daemon
empty directory	/opt/sync/emptydir	2750 confluent:confluent	700 root:root	2750 root:root

Directories the sync only has to traverse to reach those files. Nothing in the syncfile list refers to them, and
both before and after they must come out exactly as the node had them:

directory	node state before sync	before	after
parent directory (/opt/sync)	700 root:root	700 root:root	700 root:root
parent subdirectory (/opt/sync/sub)	750 daemon:daemon	750 daemon:daemon	750 daemon:daemon

The second table is the point: files were written into the subdirectory, yet neither its mode nor its ownership
moved. In the first table, the system user file keeping daemon:daemon while the confluent owned file became
root:root confirms the ownership mapping is selective.