Save a file of any size (up to 5 GB, within the member's storage limit) that the caller holds on its own computer, in three steps by action. begin: reserves the space (refused with storage_quota_exceeded, nothing reserved, when it would not fit) and answers presigned S3 upload URLs: upload.kind single (PUT the bytes to url with headers) or multipart (PUT each partSize chunk to its url, keep each response's ETag); takes filename, size_bytes, content_type. refresh: fresh URLs for named parts of a multipart save whose first URLs lapsed (they last 15 minutes); takes id, upload_id, part_numbers; answers { id, parts: [{ partNumber, url }] }. finish: after the bytes are uploaded, assembles the parts (upload_id and parts [{ partNumber, eTag }] for a multipart upload), checks what landed against the storage limit and the file type, and stores it; takes id, upload_id, parts, folder; with abort true it cancels the save and frees the reserved space. Returns id, name, size and folder like save_file. The local Connections MCP does all three steps from path on connections_storage_save_file; use this tool only from a machine that can PUT to S3.