Global rotation inconsistent with chrome

I've got a small SVG reproduction of an issue I've seen where feOffset is used to create a drop shadow, but it turns out to be inconsistent when rotated - the global transform doesn't seem to be applied to the offset, so the positioning of the offset image is wrong.

<svg height="1000" width="1000" transform="rotate(45)" xmlns="http://www.w3.org/2000/svg">
   <rect x="400" y="100" width="200" height="100" fill="black" filter="url(#myFilter)" />
   <defs>
      <filter id="myFilter" x="0" y="0" width="800" height="600" filterUnits="userSpaceOnUse">
         <feOffset in="SourceGraphic" dy="200" result="offsetImage" />
         <feBlend in="SourceGraphic" in2="offsetImage" />
      </filter>
   </defs>
</svg>

I've done the above with a couple of different values of rotate(N), across a couple of different renderers. Image were rendered to 400px, and scaled down to 200px square in the table markdown below:

rotation	resvg	Gimp (rsvg)	Chrome
0
20
45
	Seemingly rotating each item independently then applying the offset	The offset seems correctly modified, but centre of rotation isn't	I assume this is correct - and that it is rotated around the centre of the image?

There's probably a second bug/inconsistency relating to the centre of rotation. I think all of the above are identical if transformed with rotate(N 0 0), so have different ideas of what 0,0 is.

I haven't quite worked out which bit of code should be handling such adjustments - assuming it's part of is it something that should be done by convert_offset, or does this relate to the TODO: Wrong! Does not account rotate and skew in resolve_primitive_region?

Start with the SVG reproduction in the issue and inspect convert_offset and the TODO in resolve_primitive_region about rotation and skew. Compare the renderer's output with the supplied resvg and Chrome examples for the listed rotations. Done means determining the correct transform and rotation-center behavior and covering the resulting behavior with a regression test.