Searching for an element in a heap of size N is not O(K). First, it does not make sense that the time complexity for finding one element depends on the number of elements you are trying to extract (which is what K represents). Also, there is no such thing as searching in a heap — unless you count the standard look-at-every-element search in O(N).
However, finding the largest element in a heap is O(1) by design (I am obviously assuming that it’s a max-heap, so the maximum element is at the top of the heap), and removing the largest element from a heap of size N is O(log(N)) (replace it with a leaf element, and have that leaf percolate back down the heap).
So, extracting K elements from a heap, and returning the heap of non-extracted elements, would take O(K·log(N)) time.
What happens if you extract K elements non-destructively from the heap ? You can do this by keeping a heap-of-heaps (where the value of a heap is the value of its maximum element). Initially, this heap-of-heaps contains only one element (the original heap). To extract the next maximum element, you extract the top heap, extract its top element (which is the maximum) and then reinsert the two sub-heaps back into the heap-of-heaps.
This grows the heap-of-heaps by one on every removal (remove one, add two), which means it will never hold more than K elements, and so the remove-one-add-two will take O(log(K)). Iterate this, and you get an actual O(K·log(K)) algorithm that does returns the top K elements, but is unable to return the heap of non-extracted elements.