If a node wants to leave, then the best time for it to leave, would be after a relocation, because that’s when it has made best profit. If it doesn’t want to leave, then there’s not much reason to leave then, not more than with any other system.
So if a node wants to leave, could be that that’s when it will leave - after relocation. A node that wanted to leave, would leave anyway, so it’s just happening at some other point. It could be positive that it becomes more predictable to the network.
If there is something pushing the pressure up on nodes to leave, having a bunch of accumulated rewards waiting to be claimed would act as a counter - up to a certain point.
Right now in the code, nothing, still being implemented 
What is most sane to happen?
From the top of my head I’d say that it’s most sane that within the same period that a node can reconnect to the network and present itself as the same node and take the same place in the network again (a feature which is not implemented yet) - within that period it can also claim its rewards.
If the network does not recognize/approve of it as a previously connected node, then it cannot. What happens to that accumulated reward is not very important, the simplest thing is to just reset the counter (it is not money yet…).
So, this depends on that feature, which is yet to be implemented (reconnecting to the network).
If it is proven malice that gets the node kicked out, then I’d say that it makes sense that they can say goodbye to the accumulated rewards as well.
I mention this in response to your previous questions:
Rewarding on GETs has the same problem with unpopular data: A node could drop it as to store newer (more popular, profitable) chunks.
You could increase payment for the unpopular data, but that delves into more and more convoluted code, because some data is simply not going to be requested again (practically), so giving infinite reward for it is not helping (rather it would encourage false GETs on it, which isn’t helping the system).
There is a need for checks that nodes has the data it says it has, in some way or another. So that is an assumption that the farming rests on; that we will implement that.
These are the simulations I wrote about on the forum about a year ago now. Last spring I was doing a lot of simulations, and what I saw consistently is that there needs to be a huge adoption and a substantial net minting to not stay at just a few percent farmed year after year.