eval-gates

Use when the question is whether a change made an AI system or test suite worse, or whether an eval result can be trusted — "did my prompt change make it worse", "my evals are flaky", "can I bump the threshold to get CI green", "the score came back NaN", "set up evals before we ship". ALWAYS use when a gate is red and someone wants to raise/lower a threshold, widen a noise band, downgrade a metric, refresh the baseline, or otherwise make CI pass, and when a check "passed" suspiciously. Covers fail-closed release gates, thresholds, baselines, noise bands, metric registries, BLOCK fixtures, and the failure modes naive gates miss.