Didn't get a chance to test the 12900k yet, but did test a 12400 which has no e cores, only 6 p cores (12 threads). I get the same outcome: each core runs at ~350 h/s, 4 cores at ~1000 and more than that only gets slower.
I did test the 12900k at last and the results are different. This time the best peroformance was when using 12 cores, 8p+4e (out of 8p+8e total available), however running 8p+2e is just about the same. At least this time it's using most cores...
Then I tested some other algos and got more confused:
Yespower with R=8 (tide coin) does best when using all physical cores, the more the better, both p and e. It's not "restricted" to 4 p cores on the 12400 and 12700s like generic yespower/power2b.
MinotaurX performs best on the 12400 when using all 12 threads! Not just the physical. That was a surprise for me. On 12700s best is only physical cores, 8p+4e.