It turns out at least the expanded Euclidean issue is related to the precision of clamping from round-off errors between the subtraction of the norms and the dot product (there can be an error of 1e-5, for example, which would end up becoming 0.003 once the sqrt is taken. Thus tiny errors get magnified from the sqrt.

Prior to this change, we were just taking any small value < 1e-4 and clamping it to 0. The problem is that we were limiting the effective range of inputs to the algorithms. This change adds an explicit check that the norms are equal as well. Be cautioned that in order to determine the vectors are equal, we need to test that the dot product = normX = normY. The challenge is that this roundoff error makes normX == normY but the dot product might only be approximately equal to them. Adding this extra check for the equality of the norms should at least narrow down the potential for the clamping to affect other numbers, however, it can still effect things like true orthognality, which could still have equal norms.

/merge