Disable "Apply AI suggestions" workflow action on documents with no content

Description When the "Apply AI suggestions" workflow action runs on a document whose OCR produced no text at all (document.content == "", for example a photo or a blank scan), the action does not skip the document. It still calls the LLM with an empty content block plus the "similar documents" RAG context and taxonomy candidates. With nothing to classify, the model simply reproduces the metadata of whichever neighbour the vector search returned, and the action then applies that to the document: title, correspondent, document type, created date and tags. The result is a document that looks fully classified but whose metadata belongs to a completely different document. With ai_overwrite_existing enabled this also overwrites the filename-based title and the consumption date. Because the log line says "Applied AI suggestions [...]" with no warning, there is nothing that tells the user the classification was made without any document text. In my c…