It seems to be to help the artist position the gradient precisely by decoupling start and end point from the direction of the gradient. (The code to compute an equivalent two-points version is simple.)

I can vaguely understand how this is useful for a designer, but this is not a source format, and the information is redundant for rendering. Do gradients behave differently under transformations when defined this way? I don't think so, but if that were the case I could see the point.

Thinking more: perhaps it has a similar function as the centerX and centerY values for PaintVarRotate? As in: redundant for static rendering, but indispensible when varying/animating things?

@behdad confirmed that variability is indeed a factor in this design, and showed me an enlightning use case:

• imagine a rectangle with a linear gradient from the top left point to the top right point, but skewed so the "left" gradient color end line goes towards the bottom right. Without the third point, it's hard (impossible?) to make this rectangle variable.

I'm satisfied with this explanation and I'm happy to close, unless you decide that this needs mentioning in the spec, and want to keep it open as a reminder.

See also #94 and #55.