I noticed this on my system a little while ago as well. I sent an email to Lukas (Spamhalter author) asking for clarification, his response:
"Correction is not needed at this time, because this message will be classified as you want with current database state. It do just learning (not correction) in this case."
So I think it does learn from the nospam "correction" but I'm not sure what Lukas means by "current database state". It may mean the correction has already been acknowledged with the first correction request and all similar correction requests after that are no longer needed as Spamhalter already knows to do the correction at the next refresh of the database?
I'm not convinced this is working as one would expect though, to be sure I added the sender address to the white list, seems to be ok when I do that. (not really a solution though)
If anyone has more information on this issue I'd be happy to learn more about it too.
-- s --