This is the same issue as was fixed with T39486: the adjustment pass that tries to equalize different widths at either end of an edge sometimes causes the widths to get bigger and bigger. The previous fix was to let "clamp_overlap" do double duty as a way to limit this behavior. But clearly this is undiscoverable, as the current bug report shows. So I put in an "auto-limiting" mode that detects when adjustments are going crazy and then acts as if clamp_overlap were set. The reason we can't always act as if clamp_overlap is set is that certain models (e.g., Bent_test in regression tests) look bad if that is enabled.