All things being equal, I'd start with just one full-stack test that used a typical existing instance of a spreadsheet as reference for its test data, then I'd

All things being equal, I'd start with just one full-stack test that used a typical existing instance of a spreadsheet as reference for its test data, then I'd automate a browser to fill in all the inputs into your new single-page app, and I'd finally assert that the computations match the results on the reference spreadsheet. That way, the full-stack test treats the new application as a black box and validates that the new implementation is still working like the old one (at least in the cases of whatever reference spreadsheets you throw at it).