Ref T7149. We can't compute hashes of large files efficiently, but we can resume uploads by the same author, with the same name and file size, which are only partially completed. This seems like a reasonable heuristic that is unlikely to ever misfire, even if it's a little magical.

• Forced chunking on.
• Started uploading a chunked file.
• Closed the browser window.
• Dropped it into a new window.
• Upload resumed (!!!)
• Did this again.
• Downloaded the final file, which successfully reconstructed the original file.