LLMDocumentContentExtractor treats a chat generator's own "error" field as a failed call

LLMDocumentContentExtractor decides whether the per-document LLM call failed by checking if "error" in result on the chat generator's output dict. Generator outputs are free to carry extra fields, and a wrapper-style generator that always includes an error field (for example {"replies": [...], "error": None}) makes every document land in failed_documents with extraction_error=None, even though the call succeeded and returned usable content.

This is the same internal-sentinel collision class as the LLMMetadataExtractor bug fixed in #12685, on the other extractor.

Reproducer:

from unittest.mock import patch

from haystack import Document
from haystack.components.converters.image.document_to_image import DocumentToImageContent
from haystack.components.extractors.image import LLMDocumentContentExtractor
from haystack.dataclasses.chat_message import ChatMessage, ImageContent


class ErrorKeyChatGenerator:
    """Wrapper-style generator whose output dict always includes an "error" field."""

    def run(self, messages, **kwargs):
        return {"replies": [ChatMessage.from_assistant("extracted text")], "error": None}


extractor = LLMDocumentContentExtractor(chat_generator=ErrorKeyChatGenerator())

with patch.object(DocumentToImageContent, "run") as mock_convert:
    mock_convert.return_value = {
        "image_contents": [ImageContent.from_file_path("./test/test_files/images/apple.jpg")]
    }
    doc = Document(content="", meta={"file_path": "/path/to/image.pdf"})
    result = extractor.run(documents=[doc])

print({k: len(v) for k, v in result.items()})

Output (run twice, identical both times): {'documents': 0, 'failed_documents': 1} - the document is dropped with extraction_error=None even though the generator returned valid content. Expected: the document is processed and its content becomes "extracted text".

Proposed fix: in _process_llm_results, check "replies" not in result instead of "error" in result. The internal failure dicts never carry a replies key and generator outputs always do, so the failure sentinel no longer shares a namespace with generator output fields. PR follows.