Nearest neighbor k value changes results

Description

I'm noticing some weird behavior when adjusting the k value (and per_page) from 10 to 15 with my dataset. I was expecting it to simply add 5 new items, but it's adding 4 to the tail and inserting one near the head. Shouldn't increasing k value never do this or am I misunderstanding?

def find_nearest(search_embeddings, vec_field=''):
    requests = {
        'searches': [
            {
                'collection': 'games',
                'q': '*',
                'per_page': 10,
                'vector_query': f'{vec_field}vec:({json.dumps(search_embeddings)}, k:10)'
            }
        ]
    }
    response = ts.client.multi_search.perform(requests, {})['results'][0]
    # print_game each game name in response
    for item in response['hits']:
        print(item['document']['name'])

Start by reproducing the reported behavior with the Python find_nearest example, comparing vector_query k and per_page values against the games collection. Trace the nearest-neighbor search entry point and existing vector-search tests, if present, to determine whether changing k should preserve ordering; done means the behavior is corrected or its semantics are clearly documented.