Build a paired eval gate for CI

The gate works on pairs, not scores: for every case it knows whether the baseline and the candidate passed, so it can list the cases that newly fail and test whether those outweigh the cases that newly pass. Rules are per category and written before the change. The decision that matters most is generating the synthetic results as pairs, with most cases agreeing between runs as real ones do; if the two runs were drawn independently, pairing would buy nothing and the power table would mislead.