Many "high-ASCII" characters (like German umlauts) should be transscribed (as you may have already seen in HTML). The German a-umlaut (ä), for instance, should be written as "ä". What has been very helpful with German translations in this respect, was the (X)Emacs option "Modify -> Encode Characters" (or, during actual translation, the other way around: "Modify -> Decode Character Entities"). It may be worth a try with other encoding systems than latin-1 as well.
At this time, it's not clear what the practical consequences of the upcoming Unicode support in KDE 2 will be in this respect.
Screenshots and other pictures have to be in PNG format, not in GIF (for licensing reasons). Please make sure that all shots match up with all your translations of POs as well as the documentation. It might be a good idea to produce them as the very last thing before release and to permit the person in charge to make all necessary changes to POs and documentation. And again: think of a system of mutual proof-reading for all the translation work in your team. It will be much better for the overall quality of your work.
If there's a GNU Public License (GPL) contained in the documentation, please replace it with a link to the directory containing the GPL for all docs: "This program is published according to the <ulink URL="common/gpl-license.html">GNU Public LicenseL</ulink>."
The jury is still out on the question how to handle bad originals -- badly written ones, outdated ones, or documentations with factual errors. Some people say you should translate what's there, no matter what and just write an email to the author of the original. Others say, you should either correct obvious errors or leave them out while writing that email. Choose the policy yourself for your language team. But make sure that everybody sticks to it. And no matter what you do: don't just leave things as they are. Please make sure that users will get a consistent and correct GUI and online help (including consistent screenshots!) on release time.
Finally, there's the question of "IDs" in DocBook files (e.g. <sect1 ID="xyz">) which serve as HTML file names after running "db2html" -- should these be translated or not? The "official policy" regarding such section IDs is now that you can translate them. What you should not translate is whatever you find tagged with <anchor> in your texts. Otherwise, the online help will not be able to jump to certain points of the documentation.
The complete context of this problem can be seen in a posting by Eric Bischoff to the kde-i18n-doc mailing list (and the preceding thread).