Root-cause analysis from live MonkeyOCR probes: upload label is not detected, output carries no content, and deterministic bad-input 500s are treated as transient. Proposes parse_document with content-sniffed upload type and middle.json-only extraction.
