Download a Word template's body

Send `multipart/form-data` with a `file` part, JSON `{ "content_base64": "..." }`, or the raw bytes. Raw bytes are accepted under any content type that is not multipart or JSON, including the real DOCX type (`application/vnd.openxmlformats-officedocument.wordprocessingml.document`), which is what `curl --data-binary` sends. The file is parsed before it is stored, so a .docx the renderer cannot read is rejected with 422 rather than becoming a template that fails at render time. The response reports which fields the upload `added` and `removed`, so a replacement that quietly drops a variable is visible. `version` is the TEMPLATE version this upload created — the number that appears in the history and that restore and Dynamic URL pinning take. `docx_version` counts .docx bodies and is a different, lower number; do not feed it to restore. Each upload is kept under its own immutable key, so earlier versions survive. **Maximum 4.5MB on this endpoint.** It runs as a serverless function, and the platform rejects larger request bodies before this code sees them. For anything bigger — or from an agent that would otherwise have to encode the file itself — use `POST /v1/uploads`, which returns a URL the file is posted to directly and allows 25MB.