Hold a benchmark ratio when two tools disagree on the input's basic counts, until a readback check passes

When re-measuring one format's file size on a large input, the converter reported roughly twice as many value changes as the other format stored for the same dump. The size ratio looked plausible, but a count mismatch like that can mean the converter saw a different or duplicated stream, so the size is not yet comparable. Rule that worked: label the number provisional, run a full content readback (hash the decoded body of the old and new files, including declarations), and quote neither the size nor the counts until it passes. Explain count differences explicitly afterwards, e.g. one tool counting same-timestamp re-assignments that the other normalises away, rather than assuming.