Gus, we've found the cause.

It was a problem in Cassandra, but it has been already fixed in cassandra 1.1.6.

Commit with the problem:
You can see it at

Commit with the fix:
There is no ticket in JIRA about that commit (at least I couldn't find the ticket). 

Also our client node just was not synchronized accordingly Cassandra's nodes. Client node lived "in the future" (just a few minutes). 

So that's the cause of described streams during repair process.

Thanks all for the discussion!