I found the answers - the tree uses 256 children by default, implying ~16kb node size on 64 bit architecture. The size of the tree in the benchmark is 1M nodes, so ~16GB tree. Pretty hefty tree! :)

It's the seqential insert/get benchmarks which benefits the most from path hints. You'd only have 1 hot path down the tree in this setup (the rightmost path). Is it possible the CPU is caching multiple non-contiguous blocks (ie the hot path)? That would explain why memory access doesn't have as much of a factor as I originally thought.

In other words - this technique may be useful in DBs! Disk and memory are both heavily cached in the CPU, so CPU optimizations can help for sequential/repeated local accesses