Fix nested aggregation top_hits with query inner_hits

[mayya-sharipova]

Currently top_hits aggs inside nested aggs inherit the parent
query's inner_hits context, causing an error "Error retrieving path".
This happenes because nested top_hits is already operating on nested child docs,
and asking it to retrieve inner_hits adds an extra incorrect level of nestedness.

This fix proposes to create an empty InnerHitsContext for this case.

Closes #136893

[elasticsearchmachine]

Hi @mayya-sharipova, I've created a changelog YAML for you.

[elasticsearchmachine]

Pinging @elastic/es-search-relevance (Team:Search Relevance)

[elasticsearchmachine]

Pinging @elastic/es-analytical-engine (Team:Analytics)

[benwtrent]

@mayya-sharipova maybe I misunderstand how this will look for users.

But as a user, I would still expect to get inner hits for both whenever I request them. One that is present within _source the other that is present within the aggregator results. Especially since they will likely be different documents (top-hits under a terms agg will definitely return different inner hits than those that are present in the most relevant documents given the top level query).

[mayya-sharipova]

@benwtrent Thanks for your comment. This fix applies only when top_hits agg is a part of nested. When we do nested top_hits agg we already retrieving nested child docs.
So we discussed this offline:

• Documentation how nested agg works with top_hits agg
• We should first check if there is second level of nestedness and allow inner_hits, if not it makes sense to remove them.

[mayya-sharipova]

@benwtrent I've checked an it looks like inner_hits for double nested fields in top_hits doesn't work (for most docs doesn't return anything).
So something like this doesn't work:

PUT comments
{
    "mappings": {
        "properties": {
            "product": {
                "type": "keyword"
            },
            "comments": {
                "type": "nested",
                "properties": {
                  "comment": {
                    "type": "text"
                    },
                    "user": {
                      "type" : "keyword"
                    },
                    "ratings": {
                      "type": "nested",
                        "properties": {
                            "rating": {
                                "type": "integer"
                            },
                            "user": {
                              "type": "keyword"
                            }
                        }
                    }
                }
            }
        }
    }
}

POST /comments/_search?_source=false
{
  "aggs": {
    "nested_ratings": {
      "nested": {
        "path": "comments"  
      },
      "aggs": {
        "ratings_top_hits": {
            "top_hits": {
              "size": 10
            }
        }
      }
    }
  },
  "query": {
      "nested": {
          "inner_hits": {},
          "path": "comments",
          "query": {
              "match_all": {}
          }
      }
  }
}

[mayya-sharipova]

@benwtrent I want to review this PR, and appreciate your follow up review. Here is the summary:

Currently this request fails:

POST test/_search
  {
    "query": {
      "nested": {
        "inner_hits": {},           // ← Operates at parent doc scope
        "path": "nested_item",
        "query": { "match_all": {} }
      }
    },
    "aggregations": {
      "nested_item": {
        "nested": { "path": "nested_item" },  // ← Already in nested scope
        "aggregations": {
          "01": {
            "top_hits": {}          // ← Tries to inherit inner_hits, causing conflict
          }
        }
      }
    }
  }

Why it fails:

1. The query-level inner_hits operates at the parent document scope - it fetches nested children from parent documents
2. The top_hits aggregation inside nested aggregation already operates at nested document scope - it's working with nested child documents
3. When top_hits aggregation inherits the query's inner_hits context, it tries to fetch nested documents from already-nested documents (double nesting)
4. This causes "Error retrieving path" failures because the fetch phase gets confused about document offsets and contexts

The error: The system tries to extract nested documents by offset, but the offset calculation is wrong because it's applying nested logic to documents that are already in nested context.

---

This PR addresses by preventing top_hits aggregation from inheriting inner_hits when inside a nested aggregation. As a result the top_hits aggregation works correctly because it's not trying to apply double-nested logic.

---

In our last discussion we wanted to check for second level of nestedness (If the nested field has another level of nested fields inside it, then inner_hits could be useful in top_hits). I did a check, and it is currently NOT supported in Elasticsearch : double-nestedness inner_hits don't work in top_hits aggregation, producing some random results.
I suggest leave this work for future ( I can create a follow up issue) , and merge this PR to address immediate bug. WDYT?

[benwtrent]

I think this is good. We obviously cannot allow multi-layered of nested ness.