You can fix this scenario by incrementing the next_token when you commit the critical section instead of when you enter (if that's possible for your domain). This essentially makes it an optimistic lock instead of a pessimistic lock. Two threads can get the same fence token, but only one will be able to commit.

You can often rephrase and algorithm without a commit step into one that has one, although sometimes it can feel more complicated. E.g. instead of doing A and B in a critical section, commit an intent to do A and B, and then let an idempotent process execute that intent.