I think you're going for something much more complicated than what I had in mind. My idea is simply to have two modes other than accum buffer -> 8 bit alpha mask. One would be accum buffer + constant RGBA color -> update RGBA buffer. (By update I mean read an RGBA pixel, do the compositing, and write the composited pixel in place). The other would be accum buffer + source RGBA buffer -> update RGBA buffer. Of course, it's possible to imagine interleaving even more operations in the generation of the source RGBA buffer, but at some point the register pressure overcomes the r/w bandwidth.
Prefix-sum is not a particularly fast SIMD operation, due to the horizontal data dependency. I chose it for font-rs because I don't know of any faster ones that can get the job done. For just computing gradients, it's almost certainly going to be faster to compute it directly (it's simple multiply-add in the case of linear gradients) than to try to strength reduce. The same is no doubt true for SVG radial (cone) gradients.
When the gradient doesn't have any sharp creases or singularities, a very reasonable strategy is to compute it in lower resolution and then up-res, say 2x or 4x to keep the math super-simple. When it does have creases, it might make sense to decompose into regions and use different approaches in different regions.
In any case, it sounds like you may be re-inventing Cairo or Skia here. Might make sense to take a closer look at what they do and whether there's truly any low-hanging fruit left. I know that Skia has a bunch of SIMD optimizations already.
But this is fun stuff to think about and experiment with. I certainly don't want to discourage you.