The four vector tests iterated whatever the loader yielded. A loader that parses the file, misses the array and returns nothing satisfied 'every vector reproduces' perfectly -- the strongest-looking assertions in the crate were the ones that could pass having checked nothing. Rather than hard-code counts in the tests, where they rot on the next corpus change, the fixtures now declare their own: vector_count on each group, plus matched_count and refusal_count on the matching corpus. The tests assert the array length against the declared count before iterating, so corpus growth stays automatic while a mis-parse still fails. The refusal split is counted separately because that half is where the store declines to guess and asks instead; a corpus that lost every refusal vector would otherwise still read as full. Verified by emptying the vectors array with the declared count left in place: the test fails with left 0, right 65.