The threshold isn't quite 50%+. If one of the 49% nodes mines a block with a would-be censored transaction, the censoring notes have to make an economic decision to try to mine a replacement block and then a second block on top of that to rewrite the chain. At 100% this is easy, but at just 51% it's a costly gamble that will frequently fail at a huge dollar cost.
A block costs $200k at current prices. At just over 50% of the network, you are gambling at least two blocks to force a rewrite.
Absolutely agreed; however, eventually some non-hostile miner will randomly generate an acceptable `nonce` [entire point of hashing/energy-usage] which DOES include the censored-by-some transaction.
An additional function of the node-verification network (which uses essentially no energy, relative to mining) it to maintain the entirety of `mempool`, which is where unaccepted transaction-attempts live until mined into a block [which is then "accepted/denied" by same node-pool].
tl;dr: as far a probabilities go, unless you own exactly 100% of mining pool, it is impossible to censor a tx from the node's mempool; all you can do with <100% is DELAY transactions.
No single jurisdiction approaches anywhere close to majority, with #1)USA@21% †
†: https://bitnodes.io/