This is in pretty good shape I believe. If you're not in a hurry to merge, I'd like to try a possible improvement where, instead of pulling the fragments in an array first at resolve, we would iterate over the linked list (N times) and pop the closest fragment. That might sound bad at first because it's O(N²) accesses to a storage buffer (non contiguous at that).
However:

• After the first pass, the fragments might reside in cache, making accesses not so bad (to be validated of course)
• It eliminates the large array that might make the compiler allocate a ton of registers which hurts occupancy, or even spill to VRAM which is not good either.
• If that works then the number of layers is truly unlimited per pixel, the only limit being the global budget of the allocated fragment buffer.

What do you think ?

From what I see in backwards memory allocation and register-based block sort papers (they are complex and not worth it for games typically have fewer transparency layers), I think there is reason to believe in-place sorting in ssbo is slower.