Line.getSelfRect() wrong for multi-segment bezier (and degenerate symmetric segments)

[Xuepoo]

Line.getSelfRect() (and therefore getClientRect(), Transformer selection bounds, and cache bounds) is wrong for a bezier Line with more than one cubic segment — it can collapse to a degenerate zero-height/zero-width box while the actual drawn curve is much larger.

Reproduction

const line = new Konva.Line({
  bezier: true,
  // 2 + 6*2 = 14 points: a starting point + two cubic segments.
  points: [0, 50, 0, 150, 100, 150, 100, 50, 100, 50, 200, 150, 300, 50],
  stroke: 'red',
});
layer.add(line);
layer.draw();

console.log(line.getSelfRect());
// {x: 0, y: 50, width: 300, height: 0} -- height 0!

The curve itself renders correctly (dips down to about y=125 in this example); getSelfRect() just doesn't know about it.

Root cause

function getBezierExtremaPoints(points) {
  const axisPoints = [
    [points[0], points[2], points[4], points[6]],
    [points[1], points[3], points[5], points[7]],
  ];
  ...
}

This only ever reads points[0..7] — the first cubic segment — no matter how many segments the line actually has. _sceneFunc draws every points[n..n+7] group of 6 in a loop (context.bezierCurveTo(...) while n < length), so any segment after the first is invisible to the bounds calculation.

A second bug in the same function

While fixing the above I found the extrema-finding math also silently drops an axis when the cubic's derivative degenerates from quadratic to linear (a === 0, the discriminant branch's a coefficient) — which happens for a segment with symmetric control points, a fairly common shape. The if (a !== 0) guard skips that axis's extremum entirely instead of solving the resulting linear equation b*t + c = 0. This affects the original single-segment code too, not just multi-segment lines — you can reproduce it with a single symmetric segment: points: [0, 50, 0, 150, 100, 150, 100, 50].

I have a fix (rewrites getBezierExtremaPoints to iterate every segment and handles the degenerate case) + a regression test, verified against the real rendered pixels (not just the formula), and will open a PR against this issue.