One thing I've been thinking lately is that this should be baked into the process for every unit of work, not just those people expect to be difficult.

Instead of estimating by guessing based on reading the requirements, work on an issue for say an hour or so and then estimate it.

It still wouldn't be perfect, but I think this would catch a lot of cases where, for example, a seemingly simple feature ends up requiring an extensive model change.

It would also do a much better job of shaking out cases where the requirements aren't as clear as you thought when you start trying to code it.