If I understand the reporting properly all tests so far have used SipHash24:

Python 3.7.0a0 (default:5b33829badcc+, Oct 30 2016, 17:29:47) 
[GCC 4.2.1 Compatible Apple LLVM 7.3.0 (clang-703.0.31)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import sysconfig
>>> sysconfig.get_config_var("Py_HASH_ALGORITHM")
0
>>> import sys
>>> sys.hash_info.algorithm
'siphash24'

It sounds like it is worth it for me to be more rigorous and perform a battery of tests using FNV and then SipHash24 to compare:

- Performing no dispersion after the frozenset hash is initially computed from XORing entry hashes (control)
- Performing dispersion using an LCG after the frozenset hash is initially computed from XORing entry hashes (current approach)
- Performing dispersion using the selected hash algorithm (FNV/SipHash24) after the frozenset hash is initially computed from XORing entry hashes (proposed approach)

I'll take the six plots and merge them into a single PNG, and also post my (short)testing and plotting scripts for reproducibility and checking of the results.

I can also write a regression test if you think that would be good to have in the test suite (perhaps skipped by default for time), where instead of using the same seven letters a-g as test strings and varying PYTHONHASHSEED, I could perform the letter test for n=7 with 10000 different sets of short random strings to see if any fell below threshold.