Weighted random fallback flattens to uniform under high error rates

When a weighted-random fallback rejects samples and retries without replacement, high error rates cause low-weight models to be selected far more often than their weights suggest.