I traced the problem to BndLib_Box2dCurve::PerformBSpline where is just blatantly adds poles to the bounding box. This reminded of a post that Roman had done years ago. Romans work around doesn't seem to help you as the presence of triangulation doesn't seem to affect the 2d bounding box calculation. I would assume Roman is correct in speculating that this kind of 'just use the poles, hack' is for performance and any changes to a 'minimum bounding box' would slow things down.

Looking at BRepTools::OuterWire: It assigns the first wire and then iterates over the rest of the wires requiring larger boundaries in all directions. So in your case, the first wire is one of the circles that has an inflated v range. The rectangular wires v boundaries are inside of that first circular wire, causing the algo to fail. I think this algo could at least have a smarter first guess using the diagonal of the 2d bounding boxes, without much performance impact, but there would still be cases that will fail as long as the spline bounds are 'inflated'.