Andreas, Philippe, many thanks for the explanation and clarification,  
it helps me a lot.

I think leadingChar/languageTag is doubtful. Apparently it could be  
used to classify languages somewhere in the image. Nevermind.

Based on your explanations I considered to trim languageTag for now.

So I made UTF8PlainUnicodeTextConverter which trims languageTag on  
input.

UTF8TextConverter subclass: #UTF8PlainUnicodeTextConverter
nextFromStream: aStream
	|ch|
	ch := super nextFromStream: aStream.
	ch isNil ifTrue: [^nil].
	^Character value: ch asUnicode.

And I've loaded my domain objects with help of it's converter from file.
May be it's not most right solution for today but at least it works now.

Great!