[redbluevn]

Two edge types, the same mistake, and both are documented against themselves a few lines above the fault.

Spline edge

/// Position values are only X and Y, Z represents the weight.
public List<XYZ> ControlPoints { get; private set; }
...
public override BoundingBox GetBoundingBox()
{
    return BoundingBox.FromPoints(this.ControlPoints);
}

So the box's Z range is a range of weights.

Polyline edge

/// The vertex bulge is stored in the Z component.
public List<XYZ> Vertices { get; private set; }
...
public override BoundingBox GetBoundingBox()
{
    return BoundingBox.FromPoints(this.Vertices);
}

So the box's Z range is a range of bulges.

A hatch boundary is planar; both boxes are flat. And a hatch lives inside a block as often as not, so the false height propagates: the block inherits it, and every INSERT of that block inherits it again.

The polyline edge also lost real geometry

A bulged segment bows outside its own chord. The bulges were not merely misplaced into Z — they were never used as geometry at all, so the box was too small in X and Y as well. Each bulged segment is now measured through Arc, which is how the rest of the library measures an arc.

A repeated vertex that still carries a bulge is skipped: it leaves no chord for an arc to span, and Arc refuses the zero radius its own maths then produces (Circle.Radius's setter throws).

Evidence

Measured over eighteen architectural drawings: 17,315 polyline boundary edges, of which 2,302 carry a bulge and 2,106 were reporting a false Z range. After the fix, zero.

The drawings' computed extents still agree with AutoCAD's on all eighteen, so the larger X and Y boxes are geometry AutoCAD had already accounted for — the fix does not overshoot.

Tests

HatchSplineBoundingBoxTests (new file) pins the spline case, including that the box is unchanged whatever the weights are. HatchTests.APolylineBoundaryIsBoxedFlatAndFollowsItsBulge pins the polyline case: a bulge of 1 over a chord of 10 is a semicircle of radius 5, so the box must be flat in Z and reach 5 off the chord — taken from the two vertices alone it would be flat in Y too.

dotnet test: 2319 passed / 17 failed, against master at 592d70a on this machine at 2313 / 17 — the same pre-existing failures, plus the tests added here.

Related

#1223-#1230, DomCR/CSUtilities#23 — all found by comparing computed extents against the extents AutoCAD reports.