zpretty silently drops document body (data loss)

[gronke]

Tool: zpretty 4.0.2
Environment: Python 3.13, Linux. Reproduced via zpretty -i and zpretty <file> (stdout).
Severity: High - silent data loss. zpretty exits 0 and produces well-formed but truncated output; no warning is printed.

Description

Running zpretty on source/client_info.xml discards the entire <client> root element and all of its children, emitting only the XML declaration and the leading (prolog) comment:

$ zpretty source/client_info.xml
<?xml version="1.0" encoding="utf-8"?>
<!-- This file contains all known information for this client.

        All elements are MANDATORY. ...

        Example <invoice_rep></invoice_rep> -->

The <client> element (full_name, rates, everything) is gone. With zpretty -i this overwrites the file in place, so the data is lost. xmllint confirms the original is valid, well-formed XML.

The committed file source/client_info.xml reproduces it 100% of the time. Shape of the trigger:

<?xml version="1.0" encoding="UTF-8"?>
<!-- a multi-line comment in the prolog
     (before the root element) -->
<client ...>
  <full_name>X</full_name>
  <!-- inline body comments -->
  ...
</client>

Observations / partial isolation

• Removing the prolog comment (the comment between <?xml?> and <client>)
  -> output is correct.
• Keeping the prolog comment but removing all body comments
  -> output is correct.
• So the trigger is an interaction between a prolog (pre-root) comment and comments inside the body. It did not reduce to a small synthetic snippet (single-/multi-line prolog comments plus a couple of body comments format fine), so it appears sensitive to the specific document content/size. The real ~40-line file is the smallest reliable reproducer we have.

Impact

The CI XML-lint format phase (zpretty --check) cannot pass source/client_info.xml: the only "formatted" output zpretty can produce for it is the truncated one. Accepting it would mean committing data loss.

Workarounds

• Move the descriptive comment from the prolog to inside <client> - verified to make zpretty preserve the document.

Recommendation

• collective/zpretty: a formatter must never silently drop content.
• At minimum it should error out instead of emitting truncated output.