Merged cells in the left column and the row count comes back one short, silently

n=48 tables off the dunnock set, scanned schedules of benefits, one table to a page. 19 of the 48 have at least one vertically merged cell in the left colum, which is how these forms say that one benefit covers three treatment bands. Row count comes back wrong on 12 of the 19, so 63.2%. Most of them, once a merge is anywhere on the page. Two shapes to it and the second is the one that costs me. Either the merged value lands on the top row and the rows under it come back null, which i can see and can patch downstream, or it quietly folds the spanned rows into one and hands back a table that is one row short. Valid JSON, every field populated, nothing flags. I found it because a total didnt add up on a page i was looking at for something else. Horizontal merges across the header are worse but there are 3 of them in the whole set, so nothing to score there. What i want is wording that gets one output row per ruled row, so >= the number of horizontal rules minus one, and inherited values marked as inherited rather than restated as though they were printed on that line. Telling it the row count up front does nothing. It agrees with the number and then hands back a different one. These are 300 dpi scans, no digital originals, before anyone suggests reading the table out of the file. Not doing a third hand count on them, it was a faff the first time.