In each case, you delete two cells for each row. In cases 1 and 2, you first call the version ("Cell", forIndexPath: indexPath) . In this case, the table view ends with two cells for each row, one completely overlapping and obscuring the other. You can see this in the view inspector, as you can change the viewing angle to see behind:

(I changed the cellForRowAtIndexPath code as follows:
override func tableView(tableView: UITableView, cellForRowAtIndexPath indexPath: NSIndexPath) -> UITableViewCell { var cell = tableView.dequeueReusableCellWithIdentifier("plainCell", forIndexPath: indexPath) cell.textLabel!.text = "First cell for row \(indexPath.row)" cell = tableView.dequeueReusableCellWithIdentifier("plainCell", forIndexPath: indexPath) cell.textLabel!.text = "Second cell for row \(indexPath.row)" print("Cell being returned is \(cell)") return cell }
to give different text labels to each cell.) In cases 3 and 4, when you first call the version ("Cell") , there is only one cell for each row in the table view.
Why different behavior? If you create your own subclass of UITableViewCell and use it in your storyboard, you can override the various methods and add print() statements to see what happens. In particular, awakeFromNib , didMoveToSuperView and deinit . What happens is that in cases 1 and 2, the first cell is created (awakeFromNib) and immediately added (didMoveToSuperView) to the supervisor, presumably as a table or in one of its subzones. In cases 3 and 4, the first cell is created, but not added to the supervisor. Instead, after some time, the cell is freed (deinit).
(Note that if the second cell is canceled using the version ("Cell", forIndexPath: indexPath) , it is also added immediately to the supervisor. However, if the second cell is deleted using the version ("Cell") , it is only added to the supervisor after return the cellForRowAtIndexPath method.)
Thus, the key difference is that the version ("Cell", forIndexPath: indexPath) causes the cell to be immediately added to the table view before even cellForRowAtIndexPath . This hints at the question / answer you are referring to, as this indicates that the selected cell will have the correct size.
After being added to the supervisor, the first cell cannot be freed because there is still a strong link to it from its supervisor. If cells are canceled using the version ("Cell") , they are not added to the supervisor, therefore, there is no strong reference to them after the cell variable is redistributed, and therefore they are freed.
Hope this all makes sense.