fix(stats): hold the Student-t tails past t-squared overflow

Assisted-by: GLM 5.3 Flash

stats/regression.go

@@ -335,14 +335,16 @@ func LinearRegression(x, y *core.Array) (*LinearRegressionResult, error) {
 // freedom, by the closed-form tail I_z(df/2, 1/2) with z = df/(df+t²).
 // The identity is used rather than 2·(1 − T_cdf(t)): near t = 0 the
 // subtraction cancels catastrophically, while the incomplete beta
-// stays accurate into the far tail where p values matter most.
+// stays accurate into the far tail where p values matter most. The
+// upper-tail helper carries it, so the asymptotic forms that hold the
+// df ≤ 2 tails past t²'s overflow serve here too: the bare closed form
+// answers a silent 0 there while the true tail is still representable.
 func twoSidedT(t float64, df int) (float64, error) {
-	z := float64(df) / (float64(df) + t*t)
-	p, err := BetaIncomplete(z, float64(df)/2, 0.5)
+	upper, err := studentTUpperTail(math.Abs(t), df)
 	if err != nil {
 		return 0, err
 	}
-	return p, nil
+	return 2 * upper, nil
 }
 
 // hasConstantColumn reports whether an (n, p) design holds a column of