Thanks for the tips! I found some discussion, which might be related to this exact issue:
github.com/bulletphysics/bullet3/issues/182
It does indeed seem that the normal of the triangle is somehow interpreted incorrectly, which explains why the engine pushes my player physics to the wrong side.
Messing with doPhysics and margins did not fully solve the issue. By setting margins of the player model and geometry to 0, shallow slopes stabilized. But if I set for example p2 = Point3(-6, 6, 4), both the box and player model go through the triangle again. I’ll report back if I can find a more permanent solution.