Table Merge Verdict

Checks a markdown table a converter produced for the damage merged cells cause, and repairs it when you can supply the merge geometry. No I/O: it takes text, not documents.

THE SILENT FAILURE. Converting a document with merged table cells produces markdown that is well formed and wrong.

A VERTICAL MERGE SHIFTS EVERY FIELD LEFT. The affected row carries FEWER pipe-delimited cells than the header:

  | Category | Item | Value |
  | Fruit  Fruit | Apple | 10 |
  | Banana | 20 |            <- two cells against a three-cell header
  | Veg | Carrot | 30 |

dict(zip(header, row)) -- the standard way to turn a markdown table into records, and what most RAG pipelines do -- then reads Banana as the Category and 20 as the Item. Every field on that row is off by one.

SPREADSHEET MERGES FABRICATE A TOKEN. XLSX merges emit the literal string NaN into the non-anchor cells. That is not a blank: it is a data token a downstream parser or a language model reads as a real value, and it appears in the middle of otherwise correct rows.

A HORIZONTAL MERGE LOSES THE VALUE ENTIRELY, leaving blanks where the spanned content was.

THE FLAGSHIP NEEDS NOTHING BUT THE MARKDOWN. A row whose cell count differs from its header is self-evident from the text alone, so the most damaging case is detectable without any access to the original document. Supply the merge ranges as well and it also finds values lost from inside a merge, and can forward-fill them back.

IT DISTINGUISHES 'CHECKED AND FOUND NOTHING' FROM 'NEVER CHECKED'. An adversarial reviewer pointed this package's own first documented example at its own captured fixture of a real conversion data loss and got a clean verdict claiming no merge values were lost -- because no merge ranges had been supplied and that check had never run. Without ranges the verdict now says so explicitly and names the argument that would check it.

And because a real document can legitimately contain the text NaN, the fabricated-token list is documented and overridable -- and a bare string passed where a list was meant is refused by name rather than silently degrading the check to individual characters.

VERIFIED: 122 tests, measured by running the suite, every mutation observed FAILING before restore.

DELIVERY: signed download of a hash-verified tarball, immediately on purchase. Permissive licence: unlimited products, unlimited clients, unlimited seats, no attribution, perpetual and irrevocable. One restriction, do not republish the source as source.