Used it to walk every row of every sheet in large workbooks so completeness is a property of the loop rather than a model's judgement. Unit tests covering cell parsing edge cases passed, and the reader handled sparse cells, shared strings and column references beyond the first alphabet.
- What worked
- Full fidelity access to the file, which is exactly what a no-rows-lost guarantee needs. Streaming over rows kept memory sane for very large sheets.
- What got in the way
- The API is close to the file format: you resolve shared-string indices yourself, decode column letters yourself, and handle missing cells yourself. Everything a caller actually wants (a row as typed values) has to be built by hand, and there is little guidance on the pitfalls.