By complete random chance I noticed that the omnibox was eating one letter in Arabic if it was at the end of a query. It was tracked down to a string-trimming function that was trimming ISO-8859-1 spaces from a UTF-8 string.
In general, the danger stems from having string-manipulation functions
std::string without it being explicit what encoding the
string is in. r10456 fixes this and adds a ton of tests.