There are a few opportunities for both code simplification and moderate memory reduction due to @fredlas removal of shared-memory stats. Off the top of my head:

• We don't need the Gauge and Counter to share any underlying data structure so fields used by only one of them can be elided in the other. They were previously shared to make it easier to do uniform block allocation in a fixed sized shared-memory block.
• I don't think we need two levels of reference counting; we should just be able to get away with 1.
• The basic implementations of the structures are split between heap_stat_data.h and stat_data_allocator_impl.h, which were previously needed for templatizing the system to share impls between heap-allocated and raw-allocated variants. Now there is only one variant so we can un-templatize and collapse some of this.

There may be more exposed after those reductions and simplifications.