xslt rule eats part of a paragraph

At work (freitag.de) a German weekly newspaper, we use lxml to filter the content user enter (any user can register and write articles).

For a web weeks we have been trying to fix a bug where a paragraph at display time (after being processed by an XSLT transformation) gets part of its text stripped.

I can reproduce the problem in both python 2 and python 3, see below the system information.

A repository with a minimal test and instructions is available here: https://github.com/gforcada/test-lxml-bug

System information

Python : sys.version_info(major=3, minor=7, micro=9, releaselevel='final', serial=0)
lxml.etree : (4, 6, 2, 0)
libxml used : (2, 9, 10)
libxml compiled : (2, 9, 10)
libxslt used : (1, 1, 34)
libxslt compiled : (1, 1, 34)