Since nobody has provided an answer so far, I'll try to summarize my research on this. If someone is able to provide a better "story", I will switch the checkmark to it.
After reading the pBFT paper from Castor/Liskov, I believe that 66,6% + 1 rule from pBFT also applies for Clique too. Hence, 50% + 1 is not enough to be sure that the transaction can not be removed from the chain after a reorg.
So far I have only an example as a proof:
In a system with 5 sealers (N = 5), I could have a dishonest sealer S3 able to divide the network into {S1,S2,S3} and {S3,S4,S5} (for instance by an DDoS attack), so that {S2,S3} are not able to communicate with {S4,S5}. Both networks will keep running their own fork as long as S3 is participating (50%+1). When he's in-turn, he could propose a valid block with a transaction TX and publish it only to {S4,S5}, while on the other side creating an alternative block with a double spend transaction TX' publishing only to {S1,S2}. After 50%+1 confirmations in {S3,S4,S5}, he could stop publishing blocks in this subnetwork and let {S1,S2,S3} get heavier. After that, he can release the blockade in communication, and the GHOST protocol would resolve the fork by selecting the heavier chain, which is the chain created by {S1,S2,S3}.
In this sense, the minimum number of confirmations will have to be: 66,6% + 1 [more precisely: N - floor((N-1)/3)] different sealers have confirmed the transaction, which is analogous to pBFT.
About "upper limit" as defined above in the question (sum of confirmation block difficulties <= N * 2). This actually makes no sense, since S3 could let the both forks run long enough to satisfy this condition and later, he could give his desired branch an advantage by halting the communication in the other branch.
