pub(crate) const HALF_INTEGER_SERIES_T: f64 = 1.0;Expand description
Largest |t| at which Reduced sums the half-integer orders from their series in
t (which loses up to e^|t| ulps) instead of the Kambe chain (whose correlated
errors the sums cancel less the smaller |t|).
Calibrated against 30-digit references: both keep 2D spherical and 1D cylindrical
sums within 1e-13 of max(|S|, 1) for |t| from 0.3 to 1; the chain alone loses
2e-12 at |t| = 1e-3 and the series alone 1e-12 at |t| = 3.