In a briefly view, I think the opportunity is not high.
T39326 comes from the float point deviation which we don't have much to do with.

Yes, it should be.
Because in the previous discussion, we just assumed that the bone vector is already normalized, but in actual use case, We will need to normalize it.

Here comes the new patch...

When I reviewed it myself, I have found something more interesting and the algorithm is now further simplified.