Document Python/external-rsync backend differences (#20)

* Document Python/external-rsync backend differences

- Note that mixing --external-rsync (which can recreate device/socket
  entries via --archive's -D) with the default Python hardlink-forward
  step is unsafe: _recursive_hardlink() only understands files,
  symlinks, and directories and errors on anything else.
- Note that --dry-run is I/O-free with --external-rsync but still
  fully copies into a throwaway staging directory with the Python
  backend, though neither leaves a lasting change.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

* Clarify external-rsync dry-run description

Copilot review feedback on PR #20: rsync's --dry-run still reads and
stats source files to compute what would change, it just avoids
writing to the destination - "performs no I/O" overstated that.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

---------

Co-authored-by: Maxim Shulga <maxim.shulga@alludo.com>
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>

README.md

@@ -223,6 +223,22 @@ backups/
   root.** In the Python backend a failed `chown` is caught and logged,
   not fatal; with `--external-rsync`, rsync itself just skips
   ownership it isn't permitted to set.
+- **Do not mix backends across runs of the same backup.** The default
+  Python backend excludes special entries (see above) and its default
+  hardlink-forward step only understands files, symlinks, and
+  directories - it raises an error on anything else. `--external-rsync`
+  can recreate device nodes and sockets (as `root`, via `--archive`'s
+  `-D`), so a snapshot it created can contain entries the Python
+  backend's hardlinker cannot carry forward into the next backup. Pick
+  one of `--external-rsync`/`--external-hardlink` or the Python
+  defaults and use it consistently for a given backup destination.
+- **`--dry-run` behaves differently per backend, though neither leaves
+  a lasting change.** With `--external-rsync`, rsync itself avoids
+  writing to the destination (it still reads and stats files to
+  compute what would change). With the Python backend, files are still
+  fully copied into a throwaway staging directory, which is then
+  deleted - correct, but slower and more I/O-heavy than the external
+  backend's dry run.
 
 ## Development