Compositing a gradient using an unbound operator will often fail in the fallback path (thus at the very least in image), as shown by radial-gradient-source.

I've been able to track down the source of this problem to a bad interaction between pixman having gradients as 0x0 images and the fallback fixup for unbound operation, which uses the size of the images to compute the area of the dest image to be cleared (see _cairo_surface_composite_fixup_unbounded in cairo-surface.c). This only happens when the composited source has EXTEND_NONE as otherwise the source is always considered to be infinite (is this correct even when the image really is 0x0?)
Image passes radial-gradient-source in wip/compositor (where the old fallback path has been removed).

Maybe this should be a blocker for 1.10, even if it's not a regression, afaict

(I know that this testsuite failure is available to all the developers, but I wanted to keep track of the observations I made by debugging and inspecting the code, since I have no patch for this bug, yet)