The OCaml people at Jane Street noted that if you just focus on cache lines, the overhead these days between list traversal vs. vector traversal can be closer to a factor of 2. Like 16 cycles vs. 30 cycles for a new cache line on an old Skylake architecture, thanks to prefetching -- explicit prefetching instructions are hard to get right, but CPUs have also gotten better in the last 10 years at automatic prefetching. (It might be even better on Apple's architecture which I believe has more line fill buffers?) A miss is still around 300 cycles though, and a cache line of 64 bytes could very well represent more than one item (e.g. 8 longs).
This is the sort of thing that I was referring to when I referred to "cache-aware implementations of Lisps".