v0.2.1

Adds the bespoke-checker baseline: the strong hand-rolled comparator a competent team writes in a sprint. On the live set it scores 18 of 21, with one false safe and two false blocks, which is the gap between comparing fields and correlating operations.

v0.2.0

Balances the scenario set (seven safe completions, fourteen traps) and adds the always-block baseline. Every run now reports false safes and false blocks separately, so a policy that blocks everything scores exactly as badly as it deserves.