Three disciplined iterations on the enquiries drafter

A claude -p harness that keys every result by a hash of the prompt file, 36 synthetic cases split 24 dev / 12 test with a fixed seed, a paired fixed/broke table per case type, three single-change iterations recorded in a changelog that records each prompt file's hash, and a promote command that runs the frozen test split once. The worked run shows one edit rejected because it broke a control case, and why a test gain of 12 cases can't confirm a dev gain of 24.