Yes, problems can be solved after some time of work. I just think that we are far a way of live transliteration (like in Chinese). But static transliterations can be test for future live transliteration.
Markup language is a good idea, but I am not sure how often would be used. Some people would not use because they don't know how to use, some people would not use because they don't like other alphabet (the second problem is very real in Serbian culture :) ). However, it would be useful if MediaWiki would support something like "don't touch marking" (for example "---"; or something like that). For the beginning it would be used for the bot parsing; but when algorithm become stable, it should be used by MediaWiki engine.
At the page [[:sr:Корисник:Millbot/извештаји]] (http://sr.wikipedia.org/wiki/%D0%9A%D0%BE%D1%80%D0%B8%D1%81%D0%BD%D0%B8%D0%B...) can be found bot's reports during it's work one and half month ago. (I stopped with transliteration because I have to make some other parts of bot's infrastructure.)
You can find that part of bot's source code at http://millosh.org/software/ltafos/pos/ (use version 0.3.2; I have some newer on my laptop, but it is not stable). Unpack pywikipediabot (http://sourceforge.net/projects/pywikipediabot/, unpack pos, go to the pos directory and link all of pywikipediabot files into pos directory. Note that pos is in something like alpha stage and you should be careful when use it. There are some documentation in English, but Serbian documentation is better (as you are Belorussian, I think you can understand documentation in Serbian).
In this moment I am working on some other bot's tasks, but I'll finish it in month or two and I'll continue to work on transliteration. Bot Millbot should be the system with IRC and pywikipediabot parts. It should monitor channels like #srrc.wikipedia and work according to informations which it gets there: When someone make some change in Cyrillic text, it should immediately implement changes into Latin text and vice versa.
At the first stage of transliteration Millbot would make changes and synchronization only according to Cyrillic texts. When it become more clever, it should work with Latin texts, too. But it is a long time task.
Also, Bosnian Wikipedia has almost the same situation as Serbian. They starts from Latin and, according to their culture (and their Main Page), they should have Cyrillic pages, too. When I implement Latin->Cyrillic transliteration for Serbian, it would be Latin->Cyrillic transliteration implementation for Bosnian, too.
And, one more thing: I think we should move our talk to pywikipediabot list.