CPU-Z is very simple codebase (not as bad as Coremark or Dhrystone, but too "nice and easy" to be used for benchmarking). It's possible it is overstimating the IPC of some cores compared to what performance they could extract with their out of order core running more typical messy code that needs good branch prediction, prefetching and overall real-world performance to not fall apart. (Skymont has decent branch prediction perhaps better than Lion Cove, but I wonder if CPU-Z could overstate its ability to use all its execution units a bit too much.)
Using single-thread geekbench for these comparisons may actually be better idea, despite all the issues the software has (like variance, MT score being worse then useless since it is outright misleading).
I asked Kocicak (alias BoggledBeagle on techpowerup forum) for additional tests (Geekbench / Cinebench). Let's be surprised if Skymont's "Pentium M moment" is confirmed.