The merge algorithm when resolving concurrent edits seems to destroy Umlauts (possibly other non-ascii-characters as well). How to reproduce:

Assume you have a page with content "asdfäöü".

User 1 then clicks on edit and waits. User 2 clicks on edit, adds a line, lets say "jkl;" and saves. User 1 then clicks on save too (although he didn't change anything), and the page's content becomes "asdf&#xFC&#xFD&#xFE" (or whatever the LATIN1-codes for Umlauts are). Seems that the merge algorithm destroyed the umlauts.