I suspect that there's a race where a connection is established to the seeder, but we haven't received it's completed piece information yet, and we already finished reading all the data we need from another connection. Probably comes up now because pending peers with the same address aren't clobbering each other since that was fixed.