To clarify my comment, before, the code block was:

      offset += sem_ir_->associated_constants().size();
      [[fallthrough]];
    case ScopeIdTypeEnum::For<AssociatedConstantId>:

Now it looks like the code block is:

      {ScopeIdTypeEnum::For<AssociatedConstantId>,
       [](const File* sem_ir) {
         return sem_ir->associated_constants().size();
       }},

Both of these seem similar to me, with a couple differences; I always find lambdas add complexity for me, but also the new code has more general framework (i.e. more to understand).

So where you say: "The logic that sums only the entities above the given id is now a common logic so no need to change it when changing the set of ids." I don't see how the two are much different: there's a code block for a given group, and additions need an addition.

I completely understand that you may find it easier to reason about, but to me the addition of a struct and complex control flow is not fundamentally easier to reason about. If you want something simpler, I'd suggest a solution that drops both the struct and lambdas.