bayes for days

test(candidates): measure the collapse at the size the sweep actually runs master

The A/B fixture ran the enemy at 5 MP, which is a third of the positions a real 8v8 has and a memo working set that fits a cache the real one does not: the same change reads 1.05 there and 1.10 at the full size. Records what the between-process run said and why its own controls refuse it.


+39 -7
2 changed files