At each nominal level, how far empirical coverage over n trials can
wander from the diagonal by chance alone, under a Binomial(n, nominal)
model of the count that lands inside its interval. plot_coverage() and
plot_tarp() shade this as the ribbon behind their curve; plot_sbc()
uses the same interval, scaled back up to a count, for its histogram's
dashed reference lines. One formula in one place means all three figures'
bands agree on what "real" departure from calibration looks like.