What happens stemming-wise in this situation?

Elastic Stack Elasticsearch
1

I have implemented a simple module where .docx (Word) files in a directory tree are parsed and each one split into multiple (overlapping) 10-line "Lucene documents", and entered into an index using the _bulk endpoint ... and then, currently, a simple query search is done.

I have implemented stemmed fields a bit in the past but I'm trying to understand a bit more about how things work. Supposing I now add a mapping thus (NB this is added to a new, as yet empty index which I've just created with a name based on a particular regex pattern, after examining the current active index used by an alias).

mappings = {
    'properties': {
        'text_content': {
            'type': 'text',
            'term_vector': 'with_positions_offsets',
            'fields': {
                'stemmed': {
                    'type': 'text',
                    'analyzer': 'english',
                    'term_vector': 'with_positions_offsets',
                }
            }
        },
    }
}
headers = {'Content-type': 'application/json'}
success, deliverable = process_json_request(f'{ES_URL}/{new_index_name}/_mapping', 'put', data=json.dumps(mappings), headers=headers)

My code for running the queries is currently pretty much the simplest one I could possibly devise:

simple_query = { 'query': 
    { 'simple_query_string' : 
        { 'query': query_string }}}
headers = { 'Content-type': 'application/json' }
success, deliverable = utilities.process_json_request(f'{ES_URL}/{ALIAS_NAME}/_search', data=json.dumps(simple_query), headers=headers)

... I'm trying to understand: with this current type of query, will it make use of the stemmed field, or the unstemmed text?

The funny thing is, if I do a (random) query string such as "linux set", it seems to be delivering results in which stemming has indeed been applied, e.g. where I can see a word such as "setting" or "sets" is included in the delivered Lucene documents.

But then I do another query string search such as "linux sets" ... and it returns another set of results, apparently also with stemming. And another set of results when I do "linux setting"... etc.

I would have thought that if the query string is subjected to stemming before the search happens that the results would be pretty much deterministic: I would expect all three searches to be done on (as stemmed) "linux set".

In previous experiments I have moved straight on to applying a particular analyzer to my query string explicitly ... but I'm trying to understand better what exactly happens in this situation. Maybe it's explained somewhere in the docs?