Embed silently fails on append

[dainiusjocas]

Describe the bug
As of 8.738.17, when a partial update appends to an array field and an embedding is calculated from that field, the synthetic tensor field is silently not updated.

To Reproduce
Steps to reproduce the behavior:

1. Here is the full application.
2. Schema

schema doc {
  document doc {
    struct chunk {
      field text type string {}
      field sentiment type int {}
    }
    field chunks type array<chunk> {
      indexing: summary
    }
  }
  # Simply take a value from the document field which was fed and embed it.
  # truncated to just 1 dense dimension for simplicity
  field embeddings type tensor<float>(offset{}, x[1]) {
    indexing {
      input chunks |
      for_each{ get_field text } |
      embed e5 |
      attribute
    }
    attribute {
        distance-metric: angular
    }
  }
  field text type array<string> {
    indexing: input chunks | for_each{ get_field text } | index
  }
}

3. Upsert a doc:

curl -v -X PUT -H "Content-Type:application/json" --data '
  {
  "create": true,
  "update": "id:doc:doc::1",
  "fields": {
    "chunks": {
      "add": [
        {"text": "three", "sentiment": 20}
      ]
    }
  }
}' \
  http://localhost:8080/document/v1/doc/doc/docid/1

4. See the doc contents:

vespa visit --field-set="[all]" | jq

Returns:

{
  "id": "id:doc:doc::1",
  "fields": {
    "chunks": [
      {
        "text": "three",
        "sentiment": 20
      }
    ],
    "text": [
      "three"
    ]
  }
}

See that embeddings field is not returned.
While after a regular PUT, all embeddings are calculated.

Expected behavior
Something along the lines:

• maybe a note in the documentation.
• feed response could signal that something was not updated, maybe a custom header
• embedder could have a mode where tensor labels are made as a hash from the content.
  But a tough problem.

Environment (please complete the following information):

• OS: MacOS, docker
• Infrastructure: self-hosted

Vespa version
8.738.17

Additional context
Demo.

[linjiaye929]

Hi maintainers, I'd like to investigate this. I'll first add a regression reproducing an append that leaves the derived embedding stale, then trace the partial-update indexing path and propose either safe recomputation or an explicit rejection/warning. I won't broaden update semantics without agreement. Which outcome do you prefer? @linjiaye929