Disclosure: Substantially drafted by GPT-6.1 Terra for Yan Wang / LensUp, then source- and fact-checked by Codex AI. Yan Wang is affiliated with LensUp. The stated results are local Node conversions of synthetic text-only DOCX packages using the unchanged convertDocx function and installed AnyDoc WASM 0.2.4. They are not production-browser, Word visual-rendering, image, arbitrary-document or security tests. No human or owner authored, reviewed or tested this draft.
A table conversion can retain every visible label in a fixture while changing how those labels are associated. That makes a merged header a useful acceptance case: decide whether the consumer needs the original span relationship, then compare the output against that decision.
The simple-table fixture produced exactly:
| Item | Status | | --- | --- | | Draft | Ready | The merged-header fixture produced exactly:
| | | | --- | --- | | Project status | | | Draft | Ready | The empty two-cell header row is part of the observed output. In the next row, Project status survives in the first cell and the second cell is empty. The following row retains Draft and Ready. A text-presence check for those three labels would therefore pass, but it would not establish that the original two-column span relationship survived. The displayed pipe-table representation does not encode that relationship.
That observation is separate from an acceptance decision. A team might accept this flattening for a use case where retaining the words and their order is sufficient and the lost relationship does not affect interpretation. Another use case might require a rewrite or manual review because the reader must know that the original header applies across both columns. Neither policy follows automatically from the fact that the conversion completed.
A proposed acceptance policy could say: accept a flattened output only when the intended column associations can be established for the receiving use; otherwise return the table for a rewrite or manual review against the source. For a repeatable test, record the expected header relationship alongside the expected text and compare both with the representation delivered to the reader. This is a proposed check, not an additional test that was executed here.
The two fixtures answer different questions. The simple output shows the observed representation for one unmerged table; the merged output exposes the missing span relation in that specific representation. They do not establish a general rule for every table or every Markdown consumer. The actual run used @firecrawl/anydoc-wasm 0.2.4 and synthetic packages, not a corpus of real Word documents.
