standard()'s final branch only takes the copy path for target.is_file() || is_new_file_path(&target). Both are false for a target that already exists but isn't a regular file — a device node, FIFO, or socket — so it falls to InstallError::InvalidTarget, whose string is hardcoded as "No such file or directory". That's just wrong when the file is right there; install /dev/null /dev/full prints "invalid target '/dev/full': No such file or directory" instead of attempting the overwrite and reporting why it actually failed.

By the time we reach this check, is_potential_directory_path has already routed actual directories into copy_files_into_dir, so target.is_dir() is guaranteed false here — broadening the condition to target.exists() && !target.is_dir() just lets any existing non-directory target reach the same remove-then-copy path a regular file already takes, matching GNU's behavior of attempting the overwrite and surfacing the real errno.

Fixes #9934. Note: the removal-failure path this now falls into (InstallError::NotPermitted) is worded "Operation not permitted" rather than GNU's "Permission denied" for this specific EACCES case — that's a preexisting, separate string issue in that error arm and I left it alone rather than bundle an unrelated fix in here.

Didn't add a /dev/full-based test since that depends on /dev permissions that vary by environment. A more portable regression test would create a FIFO in a temp dir via the test harness and assert install overwrites it into a regular file instead of erroring — happy to add if wanted.