cc @musicinmybrain does this look OK to ensure it still works on big-endian hosts as well?

Coverage decreased (-0.05%) to 77.792% when pulling 56cea7e on jorisvandenbossche:wkb-robust-nan into 024f75a on Toblerity:shapely-2.0.

Hmm, that didn't actually help for windows. Trying the approach used in pygeos then to check the unpacked bytes with isnan

cc @musicinmybrain does this look OK to ensure it still works on big-endian hosts as well?

The current code looks OK since it uses the big_endian=False option to shapely.wkb.dumps, so the resulting byte string should be little-endian regardless of the host endianness. The code also correctly uses '>' at the beginning of the struct.unpack format string to ensure the byte string is unpacked as little-endian.

I think that using math.isnan is a good approach. Except in the very rare cases where software actually uses the NaN “payload” bits to smuggle data around, there are too many ways that assuming a particular representation can go wrong.

Thanks! Yes, the final version that I pushed shouldn't be sensitive to little/big endian host anymore, since I force little_endian in the output. Of course that then also doesn't test the big endian is correct ..

The change from this PR is included in #1213, so closing this one.