I actually realized that I have to use tangents for all the edges. So I will have to fix that.
Otherwise the thickness will only be even where you have edges "floating" inside an face.

@Campbell Barton (campbellbarton) do you have time to review this?

Looking into this, some notes:

• This changes direction calculation for edges along a chain.
  
  Even though the current method of simply interpolating the directions is limited and not always ideal, it ends of reducing possible spikes and overlaps, since the edge loop moves in a single direction instead of vertices in the chain each calculating their own local direction.

• There looks to be a bug where end-points are only offset half the width (probably simple to solve).

Comparison with existing/new behavior (the top pair has this patch applied).

Of course it's possible to make examples where this patch behaves much better too, this is just an example where it doesn't work so well.

Suggested changes:

• Keep existing logic as-is for calculating dir_side for example.
• Add a dir_side_even struct member which can have the direction used in this patch (some other way to switch between different directions would be fine too).
• Also, I don't think there needs to be a dir_tangent member added. We could instead calculate shell_v3v3_mid_normalized_to_dist and store this in for example dir_side_even_scale[2], this way the logic in doEdgeSlide can be kept quite simple.

Attached example blend file,