I continued to poke at the delay with a 2FA service (pcloud) that was previously greylisted. It scored lower on BAYES_SPAM after multiple deliveries/days - despite being the same content.
In my case, Bayes was the single deciding factor putting a mail over the threshold for greylisting. Not sure if I as a user take part in training the model (by moving mail out of spam) or if it’s global.