If 0.02 is the JND of deltaEOK, how come everybody is getting results an order of magnitude smaller? Even the author himself (at https://www.keithcirkel.co.uk/too-much-color/) says they get 0.0028, but never elaborate on the significance of that result.

JND is an average. A lot of people will do a lot worse. The measure, as I understand it, is also under different test conditions, while this is a game where people are on their own and able to - for example - tilt their head trying to find the exact angle to see the difference.

Note that the source you're seeing there is the fallback implementation, which is only used if there is no instruction for FMA in the architecture you're compiling for. On AMD64, for example, the call to math.FMA will be replaced by the VFMADD231SD instruction.

In fact, on AArch64 there is no integer multiply instruction; there's only multiply-add, and if you just want a multiply you supply the zero register as the addend.