Also kill the kink logic and thus minimisation2

Kill tests that use the old interface

Since 0 * nan = nan

x/|x| is only a problem if x == 0
just divide by |x|, no matter what non-zero value it is

Contrary to what one might expect, it is the smallest representable
*positive* number.

Using -max() works because by IEEE 754, floats are symmetric!

If |u| is computed as 0, we should not divide by it (this appears to
happen even for u != 0)