dashu_float/math/

hyper.rs

//! Hyperbolic functions, built from the cancellation-free `exp_m1` / `ln_1p` primitives:
//!
//! - `sinh(x) = (exp_m1(x) - exp_m1(-x)) / 2`
//! - `cosh(x) = (exp_m1(x) + exp_m1(-x)) / 2 + 1`
//! - `tanh(x) = exp_m1(2x) / (exp_m1(2x) + 2)`
//! - `asinh(x) = sign(x) · ln_1p(|x| + x²/(sqrt(x²+1)+1))`
//! - `acosh(x) = ln_1p((x-1) + sqrt((x-1)(x+1)))`  (x ≥ 1)
//! - `atanh(x) = ln_1p(2x/(1-x)) / 2`  (|x| < 1)
//!
//! The `exp_m1` / `ln_1p` forms avoid the catastrophic cancellation that the naive
//! `(exp(x)-exp(-x))/2` and `ln(1+…)` formulas suffer for small arguments. Special
//! values follow IEEE 754: infinities are values (not errors) for the forward functions
//! and `asinh`; `acosh(x<1)` and `atanh(|x|>1)` are domain errors.

use crate::{
    ball::{ulp_mag, ulps, Ball},
    cmp::repr_cmp_same_base,
    error::{assert_limited_precision, FpError},
    exp::pow_chain_guard,
    fbig::FBig,
    math::{
        cache::{reborrow_cache, ConstCache},
        FpResult,
    },
    repr::{Context, Repr, Word},
    round::{mode, ErrorBounds},
};
use dashu_base::{Abs, Approximation::Exact, BitTest, Sign};
use dashu_int::IBig;

impl<R: ErrorBounds> Context<R> {
    /// Hyperbolic sine.
    pub fn sinh<const B: Word>(
        &self,
        x: &Repr<B>,
        mut cache: Option<&mut ConstCache>,
    ) -> FpResult<FBig<R, B>> {
        if x.is_infinite() {
            return Ok(Exact(FBig::new(Repr::infinity_with_sign(x.sign()), *self)));
        }
        assert_limited_precision(self.precision);
        if x.significand.is_zero() {
            // sinh(±0) = ±0
            return Ok(Exact(FBig::new(signed_zero_repr(x), *self)));
        }
        // sinh(x) = (exp_m1(x) - exp_m1(-x)) / 2  (cancellation-free). Both `exp_m1` come from the
        // Ball-based `exp_compute`; the subtraction/division roundings, the `exp_m1` errors *and
        // the input's own rounding* are tracked mechanically by the Ball propagation (the raw `x`
        // is passed in — `exp_compute` rounds it to the working precision and carries the error).
        // For huge |x|, `exp_m1` overflows inside the closure and propagates; sinh(±huge) = ±inf,
        // so the sign follows `x` (the propagated error carries an intermediate sign, remapped
        // below).
        // `+ pow_chain_guard`: the closure's `exp_compute` runs the `Bⁿ` powering chain, whose
        // radius slack is charged by the target precision (the in-closure `n` keeps deriving from
        // the growing work precision, as before). Sized once, outside the loop.
        let initial_guard = self.base_guard_digits::<B>()
            + 10
            + pow_chain_guard::<B>(1usize << (self.precision.bit_len() / 2));
        self.ziv(initial_guard, |guard| {
            let work = Context::<mode::HalfEven>::new(self.precision + guard);
            let n = 1usize << (work.precision.bit_len() / 2);
            let neg_x = -x.clone();
            let ep =
                work.exp_compute::<B>(x, work.precision, true, n, reborrow_cache(&mut cache))?;
            let em =
                work.exp_compute::<B>(&neg_x, work.precision, true, n, reborrow_cache(&mut cache))?;
            let wp = work.precision;
            let s2 = addsub_guarded(&ep, &em, true, wp)?;
            let d2 = s2.div_int(2, wp)?;
            Ok(d2.to_value_radius::<R>(&Context::<R>::new(wp)))
        })
        .map_err(|_| FpError::Overflow(x.sign()))
    }

    /// Hyperbolic cosine.
    pub fn cosh<const B: Word>(
        &self,
        x: &Repr<B>,
        mut cache: Option<&mut ConstCache>,
    ) -> FpResult<FBig<R, B>> {
        if x.is_infinite() {
            // cosh(±inf) = +inf
            return Ok(Exact(FBig::new(Repr::infinity(), *self)));
        }
        assert_limited_precision(self.precision);
        if x.significand.is_zero() {
            // cosh(±0) = 1
            return Ok(Exact(FBig::new(Repr::one(), *self)));
        }

        // cosh(x) = (exp_m1(x) + exp_m1(-x)) / 2 + 1 (no cancellation: same-sign sum). Both
        // `exp_m1` come from the Ball-based `exp_compute`; the sum/divide/+1 roundings and the
        // input's own rounding (folded by `exp_compute` from the raw `x`) are tracked
        // mechanically. For huge |x|, `exp_m1` overflows inside the closure and propagates;
        // cosh(±huge) = +inf (always positive).
        // `+ pow_chain_guard`: the closure's `exp_compute` runs the `Bⁿ` powering chain, whose
        // radius slack is charged by the target precision (the in-closure `n` keeps deriving from
        // the growing work precision, as before). Sized once, outside the loop.
        let initial_guard = self.base_guard_digits::<B>()
            + 10
            + pow_chain_guard::<B>(1usize << (self.precision.bit_len() / 2));
        self.ziv(initial_guard, |guard| {
            let work = Context::<mode::HalfEven>::new(self.precision + guard);
            let n = 1usize << (work.precision.bit_len() / 2);
            let neg_x = -x.clone();
            let ep =
                work.exp_compute::<B>(x, work.precision, true, n, reborrow_cache(&mut cache))?;
            let em =
                work.exp_compute::<B>(&neg_x, work.precision, true, n, reborrow_cache(&mut cache))?;
            let wp = work.precision;
            let one = Ball::exact_int(IBig::ONE, wp);
            let half_sum = addsub_guarded(&ep, &em, false, wp)?.div_int(2, wp)?;
            Ok(addsub_guarded(&half_sum, &one, false, wp)?
                .to_value_radius::<R>(&Context::<R>::new(wp)))
        })
        .map_err(|_| FpError::Overflow(Sign::Positive))
    }

    /// Simultaneously compute `sinh(x)` and `cosh(x)` (context layer). Returns
    /// `(sinh_result, cosh_result)` where each is a [`FpResult`].
    ///
    /// This is more efficient than calling [`sinh`](Context::sinh) and [`cosh`](Context::cosh)
    /// separately, since the two share the `exp_m1(±x)` sub-computations.
    pub fn sinh_cosh<const B: Word>(
        &self,
        x: &Repr<B>,
        mut cache: Option<&mut ConstCache>,
    ) -> (FpResult<FBig<R, B>>, FpResult<FBig<R, B>>) {
        if x.is_infinite() {
            return (
                Ok(Exact(FBig::new(Repr::infinity_with_sign(x.sign()), *self))),
                Ok(Exact(FBig::new(Repr::infinity(), *self))),
            );
        }
        assert_limited_precision(self.precision);
        if x.significand.is_zero() {
            return (
                Ok(Exact(FBig::new(signed_zero_repr(x), *self))),
                Ok(Exact(FBig::new(Repr::one(), *self))),
            );
        }

        // sinh = (ep - em)/2; cosh = (ep + em)/2 + 1, sharing the two `exp_m1` calls. Certified as a
        // pair via `ziv_pair` (retry while either endpoint straddles a boundary); the input's own
        // rounding is folded by `exp_compute` from the raw `x`. For huge |x|, `exp_m1` overflows
        // inside the closure and propagates to both slots; sinh(±huge) = ±inf, cosh(±huge) = +inf,
        // so each slot's overflow sign is remapped below.
        // `+ pow_chain_guard`: the closure's `exp_compute` runs the `Bⁿ` powering chain, whose
        // radius slack is charged by the target precision (the in-closure `n` keeps deriving from
        // the growing work precision, as before). Sized once, outside the loop.
        let initial_guard = self.base_guard_digits::<B>()
            + 10
            + pow_chain_guard::<B>(1usize << (self.precision.bit_len() / 2));
        let (sinh_r, cosh_r) = self.ziv_pair(initial_guard, |guard| {
            let work = Context::<mode::HalfEven>::new(self.precision + guard);
            let n = 1usize << (work.precision.bit_len() / 2);
            let neg_x = -x.clone();
            let ep =
                work.exp_compute::<B>(x, work.precision, true, n, reborrow_cache(&mut cache))?;
            let em =
                work.exp_compute::<B>(&neg_x, work.precision, true, n, reborrow_cache(&mut cache))?;
            let wp = work.precision;
            let one = Ball::exact_int(IBig::ONE, wp);
            let sinh_ball = addsub_guarded(&ep, &em, true, wp)?.div_int(2, wp)?;
            let half_sum = addsub_guarded(&ep, &em, false, wp)?.div_int(2, wp)?;
            let cosh_ball = addsub_guarded(&half_sum, &one, false, wp)?;
            let ctx = Context::<R>::new(wp);
            Ok((sinh_ball.to_value_radius::<R>(&ctx), cosh_ball.to_value_radius::<R>(&ctx)))
        });
        (
            sinh_r.map_err(|_| FpError::Overflow(x.sign())),
            cosh_r.map_err(|_| FpError::Overflow(Sign::Positive)),
        )
    }

    /// Hyperbolic sine of `x·π`, i.e. `sinh(x·π)`.
    ///
    /// # Methodology
    /// The argument ball `t = π·x` (π from the shared constant cache, both roundings tracked by
    /// the ball arithmetic) feeds the all-positive sinh Maclaurin series for `|t| ≤ 1` — where
    /// the exponential form's `exp(t) − exp(−t)` would catastrophically cancel — and the
    /// ball `exp` composition beyond (`|exp(t)| / |exp(−t)| ≥ e²` there, a benign difference).
    /// Unlike the circular ×π functions there are no rational special points (π·x is
    /// transcendental for every nonzero rational x), so the zero case is the only exact one.
    pub fn sinh_pi<const B: Word>(
        &self,
        x: &Repr<B>,
        mut cache: Option<&mut ConstCache>,
    ) -> FpResult<FBig<R, B>> {
        if x.is_infinite() {
            return Ok(Exact(FBig::new(Repr::infinity_with_sign(x.sign()), *self)));
        }
        assert_limited_precision(self.precision);
        if x.significand.is_zero() {
            // sinh(±0) = ±0
            return Ok(Exact(FBig::new(signed_zero_repr(x), *self)));
        }

        // For huge |x| the exp composition overflows inside the closure and propagates;
        // sinh(±π·huge) = ±∞, so the sign follows x.
        // `+ pow_chain_guard`: the closure's `exp_compute` runs the `Bⁿ` powering chain, whose
        // radius slack is charged by the target precision (the in-closure `n` keeps deriving from
        // the growing work precision, as before). Sized once, outside the loop.
        let initial_guard = self.base_guard_digits::<B>()
            + 10
            + pow_chain_guard::<B>(1usize << (self.precision.bit_len() / 2));
        self.ziv(initial_guard, |guard| {
            let work = Context::<mode::HalfEven>::new(self.precision + guard);
            let wp = work.precision;
            let t = crate::math::trig::pi_scaled_ball(&work, x, reborrow_cache(&mut cache))?;
            let one = Ball::exact_int(IBig::ONE, wp);
            let val = if repr_cmp_same_base::<B, true>(&t.mid, &one.mid, None).is_le() {
                work.sinh_series(&t)?
            } else {
                let ep = work.exp_ball::<B>(&t, wp, reborrow_cache(&mut cache))?;
                let em = work.exp_ball::<B>(&-t.clone(), wp, reborrow_cache(&mut cache))?;
                addsub_guarded(&ep, &em, true, wp)?.div_int(2, wp)?
            };
            Ok(val.to_value_radius::<R>(&Context::<R>::new(wp)))
        })
        .map_err(|_| FpError::Overflow(x.sign()))
    }

    /// Hyperbolic cosine of `x·π`, i.e. `cosh(x·π)`.
    ///
    /// # Methodology
    /// The same `t = π·x` ball as [`sinh_pi`](Self::sinh_pi): the all-positive cosh series for
    /// `|t| ≤ 1`, the `(exp(t) + exp(−t))/2` sum beyond (same-sign terms, no cancellation).
    pub fn cosh_pi<const B: Word>(
        &self,
        x: &Repr<B>,
        mut cache: Option<&mut ConstCache>,
    ) -> FpResult<FBig<R, B>> {
        if x.is_infinite() {
            // cosh(±inf) = +inf
            return Ok(Exact(FBig::new(Repr::infinity(), *self)));
        }
        assert_limited_precision(self.precision);
        if x.significand.is_zero() {
            // cosh(±0) = 1
            return Ok(Exact(FBig::new(Repr::one(), *self)));
        }

        // For huge |x| the exp composition overflows inside the closure and propagates;
        // cosh(±π·huge) = +∞ (always positive).
        // `+ pow_chain_guard`: the closure's `exp_compute` runs the `Bⁿ` powering chain, whose
        // radius slack is charged by the target precision (the in-closure `n` keeps deriving from
        // the growing work precision, as before). Sized once, outside the loop.
        let initial_guard = self.base_guard_digits::<B>()
            + 10
            + pow_chain_guard::<B>(1usize << (self.precision.bit_len() / 2));
        self.ziv(initial_guard, |guard| {
            let work = Context::<mode::HalfEven>::new(self.precision + guard);
            let wp = work.precision;
            let t = crate::math::trig::pi_scaled_ball(&work, x, reborrow_cache(&mut cache))?;
            let one = Ball::exact_int(IBig::ONE, wp);
            let val = if repr_cmp_same_base::<B, true>(&t.mid, &one.mid, None).is_le() {
                work.cosh_series(&t)?
            } else {
                // cosh(t) = (e^t + e^−t)/2 with the *true* exponentials (no `+1` — that belongs
                // to the exp_m1 form `cosh` uses below on the radian side).
                let ep = work.exp_ball::<B>(&t, wp, reborrow_cache(&mut cache))?;
                let em = work.exp_ball::<B>(&-t.clone(), wp, reborrow_cache(&mut cache))?;
                addsub_guarded(&ep, &em, false, wp)?.div_int(2, wp)?
            };
            Ok(val.to_value_radius::<R>(&Context::<R>::new(wp)))
        })
        .map_err(|_| FpError::Overflow(Sign::Positive))
    }

    /// Simultaneously compute `sinh(x·π)` and `cosh(x·π)` (context layer).
    ///
    /// This is more efficient than calling [`sinh_pi`](Self::sinh_pi) and
    /// [`cosh_pi`](Self::cosh_pi) separately, since the two share the argument ball and the
    /// exponential composition.
    pub fn sinh_cosh_pi<const B: Word>(
        &self,
        x: &Repr<B>,
        mut cache: Option<&mut ConstCache>,
    ) -> (FpResult<FBig<R, B>>, FpResult<FBig<R, B>>) {
        if x.is_infinite() {
            return (
                Ok(Exact(FBig::new(Repr::infinity_with_sign(x.sign()), *self))),
                Ok(Exact(FBig::new(Repr::infinity(), *self))),
            );
        }
        assert_limited_precision(self.precision);
        if x.significand.is_zero() {
            return (
                Ok(Exact(FBig::new(signed_zero_repr(x), *self))),
                Ok(Exact(FBig::new(Repr::one(), *self))),
            );
        }

        // Certified as a pair via `ziv_pair` (retry while either endpoint straddles a boundary);
        // the overflow sign is remapped per slot as in `sinh_cosh`.
        // `+ pow_chain_guard`: the closure's `exp_compute` runs the `Bⁿ` powering chain, whose
        // radius slack is charged by the target precision (the in-closure `n` keeps deriving from
        // the growing work precision, as before). Sized once, outside the loop.
        let initial_guard = self.base_guard_digits::<B>()
            + 10
            + pow_chain_guard::<B>(1usize << (self.precision.bit_len() / 2));
        let (sinh_r, cosh_r) = self.ziv_pair(initial_guard, |guard| {
            let work = Context::<mode::HalfEven>::new(self.precision + guard);
            let wp = work.precision;
            let t = crate::math::trig::pi_scaled_ball(&work, x, reborrow_cache(&mut cache))?;
            let one = Ball::exact_int(IBig::ONE, wp);
            let (sh, ch) = if repr_cmp_same_base::<B, true>(&t.mid, &one.mid, None).is_le() {
                (work.sinh_series(&t)?, work.cosh_series(&t)?)
            } else {
                // true exponentials — no `+1` (see `cosh_pi`)
                let ep = work.exp_ball::<B>(&t, wp, reborrow_cache(&mut cache))?;
                let em = work.exp_ball::<B>(&-t.clone(), wp, reborrow_cache(&mut cache))?;
                (
                    addsub_guarded(&ep, &em, true, wp)?.div_int(2, wp)?,
                    addsub_guarded(&ep, &em, false, wp)?.div_int(2, wp)?,
                )
            };
            let ctx = Context::<R>::new(wp);
            Ok((sh.to_value_radius::<R>(&ctx), ch.to_value_radius::<R>(&ctx)))
        });
        (
            sinh_r.map_err(|_| FpError::Overflow(x.sign())),
            cosh_r.map_err(|_| FpError::Overflow(Sign::Positive)),
        )
    }

    /// Near-correct `sinh` series `S(t) = t + t³/3! + t⁵/5! + …` on `|t| ≤ 1`: all terms share
    /// the sign of t, and the ratio between consecutive terms is `t²/((2k)(2k+1)) ≤ 1/6`, so
    /// the omitted tail is bounded by the first omitted term (< 1 ulp, as for the sine series).
    fn sinh_series<const B: Word>(self, x: &Ball<B>) -> Result<Ball<B>, FpError> {
        let wp = self.precision;
        let x2 = x.mul(x, wp)?;
        let mut sum = x.clone();
        let mut term = x.clone();
        let mut k = 1usize;
        loop {
            term = term.mul(&x2, wp)?.div_int((2 * k) * (2 * k + 1), wp)?;
            if term.mid_le_ulp_lb(&sum, wp) {
                break;
            }
            sum = sum.add(&term, wp)?;
            k += 1;
        }
        sum.add_error(ulps::<B>(&sum.mid, wp, 2));
        Ok(sum)
    }

    /// Near-correct `cosh` series `C(t) = 1 + t²/2! + t⁴/4! + …` on `|t| ≤ 1`
    /// (see [`sinh_series`](Self::sinh_series)).
    fn cosh_series<const B: Word>(self, x: &Ball<B>) -> Result<Ball<B>, FpError> {
        let wp = self.precision;
        let x2 = x.mul(x, wp)?;
        let one = Ball::exact_int(IBig::ONE, wp);
        let mut sum = one.clone();
        let mut term = one;
        let mut k = 1usize;
        loop {
            term = term.mul(&x2, wp)?.div_int((2 * k) * (2 * k - 1), wp)?;
            if term.mid_le_ulp_lb(&sum, wp) {
                break;
            }
            sum = sum.add(&term, wp)?;
            k += 1;
        }
        sum.add_error(ulps::<B>(&sum.mid, wp, 2));
        Ok(sum)
    }

    /// Hyperbolic tangent.
    pub fn tanh<const B: Word>(
        &self,
        x: &Repr<B>,
        mut cache: Option<&mut ConstCache>,
    ) -> FpResult<FBig<R, B>> {
        if x.is_infinite() {
            // tanh(±inf) = ±1
            let one = FBig::new(Repr::one(), *self);
            return Ok(Exact(if x.sign() == Sign::Negative {
                -one
            } else {
                one
            }));
        }
        assert_limited_precision(self.precision);
        if x.significand.is_zero() {
            // tanh(±0) = ±0
            return Ok(Exact(FBig::new(signed_zero_repr(x), *self)));
        }

        // tanh(x) = exp_m1(2x) / (exp_m1(2x) + 2). `exp_m1(2x)` comes from the Ball-based
        // `exp_compute` on the *exact* `2x = x + x` (so `exp_compute` folds the input's own
        // rounding into the radius); the division's rounding is tracked mechanically. For large
        // positive x it overflows → tanh = +1 (returned inline as an exact value); for large
        // negative x, exp_m1(2x) → -1 (finite), so tanh → -1 naturally.
        // `+ pow_chain_guard`: the closure's `exp_compute` runs the `Bⁿ` powering chain, whose
        // radius slack is charged by the target precision (the in-closure `n` keeps deriving from
        // the growing work precision, as before). Sized once, outside the loop.
        let initial_guard = self.base_guard_digits::<B>()
            + 10
            + pow_chain_guard::<B>(1usize << (self.precision.bit_len() / 2));
        self.ziv(initial_guard, |guard| {
            let work = Context::<mode::HalfEven>::new(self.precision + guard);
            let n = 1usize << (work.precision.bit_len() / 2);
            let two_x = x + x;
            match work.exp_compute::<B>(&two_x, work.precision, true, n, reborrow_cache(&mut cache))
            {
                Err(FpError::Overflow(_)) => Ok((FBig::<R, B>::ONE, FBig::<R, B>::ZERO)), // exact +1
                Ok(e) => {
                    let wp = work.precision;
                    let two = Ball::exact_int(IBig::from(2), wp);
                    Ok(e.div(&e.add(&two, wp)?, wp)?
                        .to_value_radius::<R>(&Context::<R>::new(wp)))
                }
                Err(other) => unreachable!("exp_m1 on finite input: {other:?}"),
            }
        })
    }

    /// Inverse hyperbolic sine.
    pub fn asinh<const B: Word>(
        &self,
        x: &Repr<B>,
        mut cache: Option<&mut ConstCache>,
    ) -> FpResult<FBig<R, B>> {
        if x.is_infinite() {
            return Ok(Exact(FBig::new(Repr::infinity_with_sign(x.sign()), *self)));
        }
        assert_limited_precision(self.precision);
        if x.significand.is_zero() {
            // asinh(±0) = ±0
            return Ok(Exact(FBig::new(signed_zero_repr(x), *self)));
        }

        // asinh(x) = sign(x) · ln_1p(|x| + x²/(sqrt(x²+1)+1)) — the x²/(sqrt+1) form avoids the
        // `sqrt(x²+1) − 1` cancellation near 0. The composition is tracked as a [`Ball`] from the
        // rounded input up: the input's own rounding, the sqr, sqrt, division and the `ln_1p`
        // input error all propagate mechanically (a hand-picked per-op ulp count cannot see the
        // operand errors inherited through the chain). The `|x|` so large that `x²` overflows arm
        // falls back to the asymptotic `sign·ln(2|x|)` on the exact `2|x|` (whose input rounding
        // `ln_compute` folds itself).
        // `+ pow_chain_guard`: the closure's `exp_compute` runs the `Bⁿ` powering chain, whose
        // radius slack is charged by the target precision (the in-closure `n` keeps deriving from
        // the growing work precision, as before). Sized once, outside the loop.
        let initial_guard = self.base_guard_digits::<B>()
            + 10
            + pow_chain_guard::<B>(1usize << (self.precision.bit_len() / 2));
        self.ziv(initial_guard, |guard| {
            let work = Context::<mode::HalfEven>::new(self.precision + guard);
            let sign = x.sign();
            let wp = work.precision;
            let x_ball = Ball::from_rounded(work.repr_round_ref(x), wp);
            let abs_x_ball = if sign == Sign::Negative {
                -x_ball
            } else {
                x_ball
            };
            let one = Ball::exact_int(IBig::ONE, wp);
            let res = match abs_x_ball.sqr(wp) {
                Ok(x_sq_ball) => {
                    let sqrt_plus_one = x_sq_ball.add(&one, wp)?.sqrt(wp)?.add(&one, wp)?;
                    let arg = abs_x_ball.add(&x_sq_ball.div(&sqrt_plus_one, wp)?, wp)?;
                    work.ln_1p_ball::<B>(&arg, reborrow_cache(&mut cache))
                }
                // |x| so large that x² overflows: asinh(x) ≈ sign·ln(2|x|).
                Err(FpError::Overflow(_)) => {
                    let two_abs = Repr::new((x.significand() * IBig::from(2)).abs(), x.exponent());
                    work.ln_compute::<B>(
                        &two_abs,
                        work.precision,
                        false,
                        reborrow_cache(&mut cache),
                    )
                }
                Err(other) => unreachable!("sqr: {other:?}"),
            }?;
            let result = if sign == Sign::Negative { -res } else { res };
            Ok(result.to_value_radius::<R>(&Context::<R>::new(wp)))
        })
    }

    /// Inverse hyperbolic cosine. Domain: `x ≥ 1`.
    pub fn acosh<const B: Word>(
        &self,
        x: &Repr<B>,
        mut cache: Option<&mut ConstCache>,
    ) -> FpResult<FBig<R, B>> {
        if x.is_infinite() {
            if x.sign() == Sign::Negative {
                return Err(FpError::OutOfDomain);
            }
            return Ok(Exact(FBig::new(Repr::infinity(), *self)));
        }
        assert_limited_precision(self.precision);
        // Domain x ≥ 1, compared *exactly* (no precision argument) because `1` is a vertical
        // tangent of `acosh`: `1 + 10⁻¹⁵⁵` at 100 digits rounds onto `1`, but its true value is
        // `√2·10⁻⁷⁸`, not `0`.
        if x.sign() == Sign::Negative {
            return Err(FpError::OutOfDomain);
        }
        match repr_cmp_same_base::<B, true>(x, &Repr::<B>::one(), None) {
            core::cmp::Ordering::Less => return Err(FpError::OutOfDomain),
            core::cmp::Ordering::Equal => return Ok(Exact(FBig::new(Repr::zero(), *self))),
            core::cmp::Ordering::Greater => {}
        }

        // acosh(x) = ln_1p((x-1) + sqrt((x-1)(x+1))) — the (x-1)(x+1) form avoids the `x²−1`
        // cancellation near x = 1. The composition is tracked as a [`Ball`] from the rounded input
        // up: the input's own rounding flows into `x−1`/`x+1` and through the product, sqrt,
        // addition and the `ln_1p` input error mechanically. (The input rounding is what a
        // per-op ulp count on the *result* of `x−1` misses: near x = 1 the subtraction cancels
        // the value — and its ulp — down by |x−1|, while the inherited operand error stays at
        // ulp(x); an under-estimated radius then lets Ziv certify the wrong neighbour of a tie.)
        // The `(x-1)(x+1)` overflow arm falls back to the asymptotic `ln(2x)` on the exact `2x`
        // (whose input rounding `ln_compute` folds itself).
        // `+ pow_chain_guard`: the closure's `exp_compute` runs the `Bⁿ` powering chain, whose
        // radius slack is charged by the target precision (the in-closure `n` keeps deriving from
        // the growing work precision, as before). Sized once, outside the loop.
        let initial_guard = self.base_guard_digits::<B>()
            + 10
            + pow_chain_guard::<B>(1usize << (self.precision.bit_len() / 2));
        self.ziv(initial_guard, |guard| {
            let work = Context::<mode::HalfEven>::new(self.precision + guard);
            let wp = work.precision;
            let x_ball = Ball::from_rounded(work.repr_round_ref(x), wp);
            let one = Ball::exact_int(IBig::ONE, wp);
            let xm1_ball = x_ball.sub(&one, wp)?;
            let xp1_ball = x_ball.add(&one, wp)?;
            let res = match xm1_ball.mul(&xp1_ball, wp) {
                Ok(prod_ball) => {
                    let arg = xm1_ball.add(&prod_ball.sqrt(wp)?, wp)?;
                    work.ln_1p_ball::<B>(&arg, reborrow_cache(&mut cache))
                }
                // (x-1)(x+1) overflowed: acosh(x) ≈ ln(2x).
                Err(FpError::Overflow(_)) => {
                    let two_x = Repr::new(x.significand() * IBig::from(2), x.exponent());
                    work.ln_compute::<B>(&two_x, work.precision, false, reborrow_cache(&mut cache))
                }
                Err(other) => unreachable!("mul: {other:?}"),
            }?;
            Ok(res.to_value_radius::<R>(&Context::<R>::new(wp)))
        })
    }

    /// Inverse hyperbolic tangent. Domain: `-1 < x < 1` (`x = ±1` → ±∞, `|x| > 1` is an error).
    pub fn atanh<const B: Word>(
        &self,
        x: &Repr<B>,
        mut cache: Option<&mut ConstCache>,
    ) -> FpResult<FBig<R, B>> {
        if x.is_infinite() {
            return Err(FpError::OutOfDomain);
        }
        assert_limited_precision(self.precision);
        if x.significand.is_zero() {
            // atanh(±0) = ±0
            return Ok(Exact(FBig::new(signed_zero_repr(x), *self)));
        }
        // Domain |x| < 1: |x| = 1 → ±∞ (value), |x| > 1 → error. The boundary is compared
        // *exactly* (no precision argument) because `±1` is a pole of `atanh`: `1 − 10⁻¹⁵⁵`
        // at 100 digits rounds onto `1` but is the finite `(p+155)·ln(10)/2`, not `∞`.
        match repr_cmp_same_base::<B, true>(x, &Repr::<B>::one(), None) {
            core::cmp::Ordering::Greater => return Err(FpError::OutOfDomain),
            core::cmp::Ordering::Equal => {
                return Ok(Exact(FBig::new(Repr::infinity_with_sign(x.sign()), *self)));
            }
            _ => {}
        }

        // atanh(x) = ln_1p(2x/(1-x)) / 2. The ratio and the `ln_1p` input error are tracked as a
        // [`Ball`]; near |x| = 1 the `2x/(1-x)` division amplifies, but the Ball tracks it (Ziv
        // retries there).
        // `+ pow_chain_guard`: the closure's `exp_compute` runs the `Bⁿ` powering chain, whose
        // radius slack is charged by the target precision (the in-closure `n` keeps deriving from
        // the growing work precision, as before). Sized once, outside the loop.
        let initial_guard = self.base_guard_digits::<B>()
            + 10
            + pow_chain_guard::<B>(1usize << (self.precision.bit_len() / 2));
        self.ziv(initial_guard, |guard| {
            let work = Context::<mode::HalfEven>::new(self.precision + guard);
            let wp = work.precision;
            let x_ball = Ball::from_rounded(work.repr_round_ref(x), wp);
            let one = Ball::exact_int(IBig::ONE, wp);
            let ratio = x_ball
                .scale_int(&IBig::from(2), wp)?
                .div(&one.sub(&x_ball, wp)?, wp)?;
            let res = work.ln_1p_ball::<B>(&ratio, reborrow_cache(&mut cache))?;
            Ok(res
                .div_int(2, wp)?
                .to_value_radius::<R>(&Context::<R>::new(wp)))
        })
    }
}

impl<R: ErrorBounds, const B: Word> FBig<R, B> {
    /// Calculate the hyperbolic sine of the floating point number.
    ///
    /// # Examples
    ///
    /// ```
    /// # use core::str::FromStr;
    /// # use dashu_base::ParseError;
    /// # use dashu_float::DBig;
    /// let a = DBig::from_str("0.5000000")?;
    /// assert_eq!(a.sinh(), DBig::from_str("0.52109531")?);
    /// # Ok::<(), ParseError>(())
    /// ```
    #[inline]
    pub fn sinh(&self) -> Self {
        self.context.unwrap_fp(self.context.sinh(&self.repr, None))
    }

    /// Calculate the hyperbolic cosine of the floating point number.
    ///
    /// # Examples
    ///
    /// ```
    /// # use core::str::FromStr;
    /// # use dashu_base::ParseError;
    /// # use dashu_float::DBig;
    /// let a = DBig::from_str("0.5000000")?;
    /// assert_eq!(a.cosh(), DBig::from_str("1.127626")?);
    /// # Ok::<(), ParseError>(())
    /// ```
    #[inline]
    pub fn cosh(&self) -> Self {
        self.context.unwrap_fp(self.context.cosh(&self.repr, None))
    }

    /// Simultaneously calculate the hyperbolic sine and cosine of the number.
    ///
    /// This is more efficient than calling [`sinh`](FBig::sinh) and [`cosh`](FBig::cosh)
    /// separately, since the two share the `exp_m1(±x)` sub-computations.
    ///
    /// # Examples
    ///
    /// ```
    /// # use core::str::FromStr;
    /// # use dashu_base::ParseError;
    /// # use dashu_float::DBig;
    /// let a = DBig::from_str("0.5000000")?;
    /// let (s, c) = a.sinh_cosh();
    /// assert_eq!(s, DBig::from_str("0.52109531")?);
    /// assert_eq!(c, DBig::from_str("1.127626")?);
    /// # Ok::<(), ParseError>(())
    /// ```
    #[inline]
    pub fn sinh_cosh(&self) -> (Self, Self) {
        let (s, c) = self.context.sinh_cosh(&self.repr, None);
        (self.context.unwrap_fp(s), self.context.unwrap_fp(c))
    }

    /// Calculate the hyperbolic sine of the floating point number multiplied by π, i.e.
    /// `sinh(self·π)`.
    ///
    /// Unlike [`sinh`](FBig::sinh), the ×π variant's argument ball is built from the shared π
    /// constant, so repeated calls at increasing precision reuse the cached Chudnovsky state.
    ///
    /// # Examples
    /// ```
    /// # use core::str::FromStr;
    /// # use dashu_base::ParseError;
    /// # use dashu_float::DBig;
    /// // sinh(π/2) ≈ 2.30129890
    /// let a = DBig::from_str("0.50000000")?;
    /// assert_eq!(a.sinh_pi(), DBig::from_str("2.30129890")?);
    /// # Ok::<(), ParseError>(())
    /// ```
    #[inline]
    pub fn sinh_pi(&self) -> Self {
        self.context
            .unwrap_fp(self.context.sinh_pi(&self.repr, None))
    }

    /// Calculate the hyperbolic cosine of the floating point number multiplied by π, i.e.
    /// `cosh(self·π)`.
    ///
    /// # Examples
    /// ```
    /// # use core::str::FromStr;
    /// # use dashu_base::ParseError;
    /// # use dashu_float::DBig;
    /// // cosh(π/2) ≈ 2.50917848
    /// let a = DBig::from_str("0.50000000")?;
    /// assert_eq!(a.cosh_pi(), DBig::from_str("2.50917848")?);
    /// # Ok::<(), ParseError>(())
    /// ```
    #[inline]
    pub fn cosh_pi(&self) -> Self {
        self.context
            .unwrap_fp(self.context.cosh_pi(&self.repr, None))
    }

    /// Simultaneously calculate the hyperbolic sine and cosine of the number multiplied by π.
    ///
    /// This is more efficient than calling [`sinh_pi`](FBig::sinh_pi) and
    /// [`cosh_pi`](FBig::cosh_pi) separately, since the two share the argument ball and the
    /// exponential composition.
    ///
    /// # Examples
    /// ```
    /// # use core::str::FromStr;
    /// # use dashu_base::ParseError;
    /// # use dashu_float::DBig;
    /// let a = DBig::from_str("0.50000000")?;
    /// let (s, c) = a.sinh_cosh_pi();
    /// assert_eq!(s, DBig::from_str("2.30129890")?);
    /// assert_eq!(c, DBig::from_str("2.50917848")?);
    /// # Ok::<(), ParseError>(())
    /// ```
    #[inline]
    pub fn sinh_cosh_pi(&self) -> (Self, Self) {
        let (s, c) = self.context.sinh_cosh_pi(&self.repr, None);
        (self.context.unwrap_fp(s), self.context.unwrap_fp(c))
    }

    /// Calculate the hyperbolic tangent of the floating point number.
    ///
    /// # Examples
    ///
    /// ```
    /// # use core::str::FromStr;
    /// # use dashu_base::ParseError;
    /// # use dashu_float::DBig;
    /// let a = DBig::from_str("0.5000000")?;
    /// assert_eq!(a.tanh(), DBig::from_str("0.46211716")?);
    /// # Ok::<(), ParseError>(())
    /// ```
    #[inline]
    pub fn tanh(&self) -> Self {
        self.context.unwrap_fp(self.context.tanh(&self.repr, None))
    }

    /// Calculate the inverse hyperbolic sine of the floating point number.
    ///
    /// # Examples
    ///
    /// ```
    /// # use core::str::FromStr;
    /// # use dashu_base::ParseError;
    /// # use dashu_float::DBig;
    /// let a = DBig::from_str("0.5000000")?;
    /// assert_eq!(a.asinh(), DBig::from_str("0.48121183")?);
    /// # Ok::<(), ParseError>(())
    /// ```
    #[inline]
    pub fn asinh(&self) -> Self {
        self.context.unwrap_fp(self.context.asinh(&self.repr, None))
    }

    /// Calculate the inverse hyperbolic cosine of the floating point number.
    ///
    /// # Panics
    ///
    /// Panics if the number is less than 1 (out of domain).
    ///
    /// # Examples
    ///
    /// ```
    /// # use core::str::FromStr;
    /// # use dashu_base::ParseError;
    /// # use dashu_float::DBig;
    /// let a = DBig::from_str("2.000000")?;
    /// assert_eq!(a.acosh(), DBig::from_str("1.316958")?);
    /// # Ok::<(), ParseError>(())
    /// ```
    #[inline]
    pub fn acosh(&self) -> Self {
        self.context.unwrap_fp(self.context.acosh(&self.repr, None))
    }

    /// Calculate the inverse hyperbolic tangent of the floating point number.
    ///
    /// # Panics
    ///
    /// Panics if the absolute value is greater than or equal to 1 (out of domain;
    /// `|x| = 1` is infinite and `|x| > 1` is not real).
    ///
    /// # Examples
    ///
    /// ```
    /// # use core::str::FromStr;
    /// # use dashu_base::ParseError;
    /// # use dashu_float::DBig;
    /// let a = DBig::from_str("0.5000000")?;
    /// assert_eq!(a.atanh(), DBig::from_str("0.54930614")?);
    /// # Ok::<(), ParseError>(())
    /// ```
    #[inline]
    pub fn atanh(&self) -> Self {
        self.context.unwrap_fp(self.context.atanh(&self.repr, None))
    }
}

/// `±0` `Repr` carrying the sign of `x` (used by the odd hyperbolics at zero input).
fn signed_zero_repr<const B: Word>(x: &Repr<B>) -> Repr<B> {
    if x.is_neg_zero() {
        Repr::neg_zero()
    } else {
        Repr::zero()
    }
}

/// `a ± b` with a magnitude guard: when one operand sits more than `prec + 2` digits below the
/// other, it is dropped and the radius inflated by one ulp instead — a sound substitute (the
/// dropped term is < 1 ulp of the dominant one) for the plain ball add/sub, whose midpoint
/// alignment would try to materialize the astronomically large digit gap between the two
/// exponentials of `(exp(t) ± exp(−t))/2` (an out-of-memory panic for large |t|, e.g.
/// `sinh(1e14)`).
///
/// The guard is a caller-side shortcut, not the only defence — `repr_round_sum` also collapses
/// an astronomically low sticky — but it avoids the alignment entirely, which is the expensive
/// half of the operation.
fn addsub_guarded<const B: Word>(
    a: &Ball<B>,
    b: &Ball<B>,
    sub: bool,
    prec: usize,
) -> Result<Ball<B>, FpError> {
    let a_dominant = repr_cmp_same_base::<B, true>(&a.mid, &b.mid, None).is_ge();
    let (big, small) = if a_dominant { (a, b) } else { (b, a) };
    // `small ≤ big·B^−(prec+2)`, checked against an exponent-only shifted threshold (the log2
    // comparison inside `abs_cmp` short-circuits the astronomical gaps — no allocation).
    let threshold = Repr::<B>::new(
        big.mid.significand().clone(),
        big.mid
            .exponent()
            .saturating_sub(prec.saturating_add(2) as isize),
    );
    if repr_cmp_same_base::<B, true>(&small.mid, &threshold, None).is_le() {
        let mut r = if sub && !a_dominant {
            -big.clone()
        } else {
            big.clone()
        };
        // Drop the below-ulp term, replacing it with its bound: one ulp of the dominant value.
        r.add_error(ulp_mag::<B>(&r.mid, prec));
        Ok(r)
    } else if sub {
        a.sub(b, prec)
    } else {
        a.add(b, prec)
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::round::mode;
    use crate::round::Round;
    use crate::DBig;
    use alloc::format;
    use core::str::FromStr;
    use dashu_base::AbsOrd;
    use dashu_int::IBig;

    // `sinh`/`cosh` go through `unwrap_fp`, so a huge-|x| overflow saturates to the directed
    // endpoint: outward (Up) → ±∞ (cosh) / sign·∞ (sinh), inward (Zero) → the largest finite.
    #[test]
    fn test_sinh_cosh_directed_overflow() {
        let p = 53;
        let max_sig = (IBig::ONE << p) - IBig::ONE;
        let huge = FBig::<mode::HalfEven, 2>::from_parts(IBig::ONE << 63, 0)
            .with_precision(p)
            .value();

        let sinh_up = huge.clone().with_rounding::<mode::Up>().sinh();
        let sinh_zero = huge.clone().with_rounding::<mode::Zero>().sinh();
        assert!(
            sinh_up.repr().is_infinite() && sinh_up.repr().sign() == Sign::Positive,
            "sinh Up -> +∞"
        );
        assert_eq!(sinh_zero.repr().significand(), &max_sig, "sinh Zero -> largest finite");
        assert_eq!(sinh_zero.repr().exponent(), isize::MAX);

        let cosh_up = huge.clone().with_rounding::<mode::Up>().cosh();
        let cosh_zero = huge.clone().with_rounding::<mode::Zero>().cosh();
        assert!(
            cosh_up.repr().is_infinite() && cosh_up.repr().sign() == Sign::Positive,
            "cosh Up -> +∞"
        );
        assert_eq!(cosh_zero.repr().significand(), &max_sig, "cosh Zero -> largest finite");
        assert_eq!(cosh_zero.repr().exponent(), isize::MAX);

        // Negative huge: this is the case the closure's sign remap exists for — `exp_m1(−x)`
        // overflows carrying a positive sign that sinh must flip to negative (and cosh must leave
        // positive). Under `Up`, a negative overflow rounds inward (largest finite negative) while a
        // positive overflow reaches +∞.
        let neg_max_sig = -max_sig.clone();
        let sinh_neg_up = (-huge.clone()).with_rounding::<mode::Up>().sinh();
        assert_eq!(sinh_neg_up.repr().sign(), Sign::Negative, "sinh(−huge) sign");
        assert_eq!(
            sinh_neg_up.repr().significand(),
            &neg_max_sig,
            "sinh(−huge) Up -> largest finite"
        );
        assert_eq!(sinh_neg_up.repr().exponent(), isize::MAX);
        let cosh_neg_up = (-huge.clone()).with_rounding::<mode::Up>().cosh();
        assert!(
            cosh_neg_up.repr().is_infinite() && cosh_neg_up.repr().sign() == Sign::Positive,
            "cosh(−huge) Up -> +∞"
        );

        // sinh_cosh(−huge) = (largest finite negative, +∞) under Up — per-slot sign remap.
        let (sh, ch) = (-huge.clone()).with_rounding::<mode::Up>().sinh_cosh();
        assert_eq!(sh.repr().sign(), Sign::Negative, "sinh_cosh[0](−huge) sign");
        assert_eq!(
            sh.repr().significand(),
            &neg_max_sig,
            "sinh_cosh[0](−huge) Up -> largest finite"
        );
        assert!(
            ch.repr().is_infinite() && ch.repr().sign() == Sign::Positive,
            "sinh_cosh[1](−huge) -> +∞"
        );
    }

    /// A near-tie argument with more digits than the working precision: the input's own
    /// rounding (½ ulp of `x`, which survives the `x−1` cancellation untouched) must be part
    /// of the certified radius. Regression for the mis-rounding this input exhibited (issue
    /// #102): the true acosh sits 1.18e-133 above the 50-digit midpoint, and the dropped
    /// input error (~1e-62) let Ziv certify the lower neighbour.
    fn acosh_near_tie_arg() -> FBig<mode::HalfEven, 10> {
        use core::str::FromStr;
        FBig::<mode::HalfEven, 10>::from_str(
            "1.000008608520579272799119633794615751764072075741189452445358800365413956109281413683821004335915587724859246904",
        )
        .unwrap()
    }

    #[test]
    fn acosh_rounds_up_above_decimal_tie() {
        use core::str::FromStr;
        let x = acosh_near_tie_arg();
        let y = Context::<mode::HalfEven>::new(50)
            .acosh(x.repr(), None)
            .unwrap()
            .value();
        assert_eq!(
            y,
            FBig::<mode::HalfEven, 10>::from_str(
                "0.0041493392794990204344718343065778753953054547309876"
            )
            .unwrap()
        );
    }

    /// The near-tie argument under directed modes must match a high-precision oracle re-rounded
    /// under the same mode (the definition of correct rounding).
    #[test]
    fn acosh_asinh_directed_match_oracle_near_tie() {
        use core::str::FromStr;
        let x = acosh_near_tie_arg();
        let asinh_arg = FBig::<mode::HalfEven, 10>::from_str(
            "0.0041493392794990204344718343065778753953054547309875500000000000000000000000000000000000000000000000000000001179",
        )
        .unwrap();
        for p in [16usize, 34, 50] {
            let oracle_acosh = Context::<mode::HalfEven>::new(p + 60)
                .acosh::<10>(x.repr(), None)
                .unwrap()
                .value();
            let oracle_asinh = Context::<mode::HalfEven>::new(p + 60)
                .asinh::<10>(asinh_arg.repr(), None)
                .unwrap()
                .value();
            macro_rules! check {
                ($mode:ty, $name:expr) => {{
                    let want_acosh = Context::<$mode>::new(p)
                        .repr_round_ref(&oracle_acosh.repr)
                        .value();
                    let got_acosh = Context::<$mode>::new(p)
                        .acosh::<10>(x.repr(), None)
                        .unwrap()
                        .value();
                    assert_eq!(got_acosh.repr, want_acosh, "{} acosh p={p}", $name);
                    let want_asinh = Context::<$mode>::new(p)
                        .repr_round_ref(&oracle_asinh.repr)
                        .value();
                    let got_asinh = Context::<$mode>::new(p)
                        .asinh::<10>(asinh_arg.repr(), None)
                        .unwrap()
                        .value();
                    assert_eq!(got_asinh.repr, want_asinh, "{} asinh p={p}", $name);
                }};
            }
            check!(mode::Down, "Down");
            check!(mode::Up, "Up");
            check!(mode::Zero, "Zero");
            check!(mode::HalfEven, "HalfEven");
        }
    }

    /// Long-digit inputs through every hyperbolic inverse: the result must match a
    /// high-precision oracle under each rounding mode (the input's rounding is folded into the
    /// radius on every path — sinh/cosh/tanh through `exp_compute`, asinh/acosh/atanh through
    /// the ball compositions).
    #[test]
    fn hyper_directed_match_oracle_long_input() {
        use core::str::FromStr;
        let args = [
            "1.000008608520579272799119633794615751764072075741189452445358800365413956109281413683821004335915587724859246904",
            "12.345678901234567890123456789012345678901234567890123456789012345678901234567890123456789012345678901234567890123456",
            "-0.8608520579272799119633794615751764072075741189452445358800365413956109281413683821004335915587724859246904",
        ];
        for a in args {
            let x = FBig::<mode::HalfEven, 10>::from_str(a).unwrap();
            let p = 34;
            let ctx_h = Context::<mode::HalfEven>::new(p + 60);
            let oracle_asinh = ctx_h.asinh::<10>(x.repr(), None).unwrap().value().repr;
            let oracle_tanh = ctx_h.tanh::<10>(x.repr(), None).unwrap().value().repr;
            let oracle_sinh = ctx_h.sinh::<10>(x.repr(), None).unwrap().value().repr;
            let oracle_cosh = ctx_h.cosh::<10>(x.repr(), None).unwrap().value().repr;
            let oracle_atanh = if x.abs_cmp(&FBig::ONE).is_lt() {
                Some(ctx_h.atanh::<10>(x.repr(), None).unwrap().value().repr)
            } else {
                None
            };
            let oracle_acosh =
                if !x.repr().significand().is_zero() && x.repr().sign() == Sign::Positive {
                    Some(ctx_h.acosh::<10>(x.repr(), None).unwrap().value().repr)
                } else {
                    None
                };
            macro_rules! check {
                ($mode:ty, $name:expr, $f:ident, $oracle:expr) => {{
                    let want = Context::<$mode>::new(p).repr_round_ref(&$oracle).value();
                    let got = Context::<$mode>::new(p)
                        .$f::<10>(x.repr(), None)
                        .unwrap()
                        .value();
                    assert_eq!(got.repr, want, "{} {} p={p} x={a}", $name, stringify!($f));
                }};
            }
            check!(mode::Down, "Down", asinh, oracle_asinh);
            check!(mode::Up, "Up", asinh, oracle_asinh);
            check!(mode::Down, "Down", tanh, oracle_tanh);
            check!(mode::Up, "Up", tanh, oracle_tanh);
            check!(mode::Down, "Down", sinh, oracle_sinh);
            check!(mode::Up, "Up", sinh, oracle_sinh);
            check!(mode::Down, "Down", cosh, oracle_cosh);
            check!(mode::Up, "Up", cosh, oracle_cosh);
            if let Some(o) = &oracle_atanh {
                check!(mode::Down, "Down", atanh, *o);
                check!(mode::Up, "Up", atanh, *o);
            }
            if let Some(o) = &oracle_acosh {
                check!(mode::Down, "Down", acosh, *o);
                check!(mode::Up, "Up", acosh, *o);
            }
        }
    }

    /// Re-round a high-precision `HalfEven` oracle to precision `p` under the mode under test.
    fn reround<R: Round, const B: Word>(hi: &FBig<mode::HalfEven, B>, p: usize) -> FBig<R, B> {
        let ctx = Context::<R>::new(p);
        FBig::new(ctx.repr_round_ref(hi.repr()).value(), ctx)
    }

    /// `sinh_pi`/`cosh_pi` vs the high-precision oracle (p + 60 under `HalfEven`, re-rounded
    /// under the mode under test), across the four canonical significand widths. The inputs
    /// straddle the `|πx| = 1` series/exp boundary from both sides (0.3 → series, 0.5 → exp —
    /// where an early version wrongly carried the exp_m1 form's `+1` into the true-exp
    /// composition, returning cosh_pi(0.5) = 3.509… instead of 2.509…).
    #[test]
    fn test_sinh_cosh_pi_oracle() {
        let inputs = [
            "0.1", "0.2", "0.3", "0.5", "0.7", "1.5", "2.808", "10.1", "123.456", "-0.3", "-1.2",
            "-2.808",
        ];
        for &p in &[20usize, 50, 100, 500] {
            for input in inputs {
                let x = DBig::from_str(input).unwrap();
                let hi_ctx = Context::<mode::HalfEven>::new(p + 60);
                let hi_s = hi_ctx.sinh_pi::<10>(x.repr(), None).unwrap().value();
                let hi_c = hi_ctx.cosh_pi::<10>(x.repr(), None).unwrap().value();

                let got = Context::<mode::HalfEven>::new(p)
                    .sinh_pi::<10>(x.repr(), None)
                    .unwrap()
                    .value();
                assert_eq!(
                    got,
                    reround::<mode::HalfEven, 10>(&hi_s, p),
                    "sinh_pi({input}) at p={p}"
                );
                let got = Context::<mode::Down>::new(p)
                    .cosh_pi::<10>(x.repr(), None)
                    .unwrap()
                    .value();
                assert_eq!(
                    got,
                    reround::<mode::Down, 10>(&hi_c, p),
                    "cosh_pi({input}) at p={p} under Down"
                );

                let (s, c) = Context::<mode::HalfEven>::new(p).sinh_cosh_pi::<10>(x.repr(), None);
                assert_eq!(
                    s.unwrap().value(),
                    reround::<mode::HalfEven, 10>(&hi_s, p),
                    "sinh_cosh_pi sinh({input}) at p={p}"
                );
                assert_eq!(
                    c.unwrap().value(),
                    reround::<mode::HalfEven, 10>(&hi_c, p),
                    "sinh_cosh_pi cosh({input}) at p={p}"
                );
            }
        }
    }

    /// The two exponentials of `sinh(1e14)` sit ~10¹⁵ digits apart: the ball compositions must
    /// drop the negligible one instead of trying to align that gap (an out-of-memory panic
    /// before the guarded add existed), and the plain `sinh`/`cosh` must survive the same
    /// argument through the core sticky-collapse in the aligned sum.
    // The ~10¹⁴-scale argument needs the 64-bit `isize` exponent range; on 32-bit targets the
    // overflow guard fires first (its range is ~4000× smaller), so the test is 64-bit-only.
    #[cfg(target_pointer_width = "64")]
    #[test]
    fn test_sinh_huge_argument_no_oom() {
        let x = DBig::from_str("189870321059487.19").unwrap();
        let ctx = Context::<mode::HalfEven>::new(20);
        let v = ctx.sinh::<10>(x.repr(), None).unwrap().value();
        assert_eq!(v.repr().sign(), Sign::Positive);
        let v = ctx.sinh_pi::<10>(x.repr(), None).unwrap().value();
        assert_eq!(v.repr().sign(), Sign::Positive);
        let v = ctx.cosh::<10>(x.repr(), None).unwrap().value();
        assert_eq!(v.repr().sign(), Sign::Positive);
        let v = ctx.cosh_pi::<10>(x.repr(), None).unwrap().value();
        assert_eq!(v.repr().sign(), Sign::Positive);
        let (s, c) = ctx.sinh_cosh_pi::<10>(x.repr(), None);
        assert_eq!(s.unwrap().value().repr().sign(), Sign::Positive);
        assert_eq!(c.unwrap().value().repr().sign(), Sign::Positive);
        // odd: the negative side flips sinh's sign only
        let neg = -x;
        assert_eq!(
            ctx.sinh_pi::<10>(neg.repr(), None)
                .unwrap()
                .value()
                .repr()
                .sign(),
            Sign::Negative
        );
        assert_eq!(
            ctx.cosh_pi::<10>(neg.repr(), None)
                .unwrap()
                .value()
                .repr()
                .sign(),
            Sign::Positive
        );
    }

    // The closure's `exp_compute` runs the `Bⁿ` powering chain, whose radius slack used to cost
    // one systematic Ziv retry on non-power-of-two bases (`pow_chain_guard` now charges the chain
    // length to the initial guard). Pins a measured retrying point (DBig `sinh 10` @150 and
    // @600 both retried exactly once) to first-attempt certification.
    #[cfg(feature = "std")]
    #[test]
    fn sinh_certifies_first_attempt_base10() {
        let ctx = Context::<mode::HalfEven>::new(150);
        crate::ziv_retries_reset();
        let _ = ctx
            .sinh::<10>(&Repr::<10>::new(10.into(), 0), None)
            .unwrap();
        assert_eq!(
            crate::ziv_retries(),
            0,
            "DBig sinh(10) @150 should certify on the first attempt"
        );
    }

    // A result in `10^(0.301·isize::MAX)..10^isize::MAX` is representable as a `Repr` but lies
    // past the old `Mag` ceiling (`2^isize::MAX ≈ 10^(0.301·isize::MAX)`): the exported radius
    // saturated to `+∞`, which no Ziv attempt can certify, and the loop spun through its
    // exponential guard growth for hours before erroring. The `Mag` exponent field is now
    // `i128` — wide enough to bound any representable value — so the ×π hyperbolic family
    // certifies here in one retry. The input magnitude scales with the target's pointer width
    // (the 64-bit magnitude is a genuine overflow on a 32-bit `isize`).
    #[cfg(feature = "std")]
    #[test]
    fn sinh_pi_huge_magnitude_certifies() {
        let ctx = Context::<mode::HalfEven>::new(20);
        // aim `x` at `0.4·isize::MAX`: the result's decimal exponent is `π·x/ln10 ≈ 0.546·MAX`,
        // comfortably past the old ceiling (`0.301·MAX`) and short of the `Repr` range (`MAX`),
        // on every pointer width (the 64-bit magnitude would be a genuine overflow on a
        // 32-bit `isize`)
        let imax = isize::MAX as f64;
        let e = (0.4 * imax).log10().floor() as isize;
        let sig = (0.4 * imax / 10f64.powi(e as i32)).round() as i64;
        let x = Repr::<10>::new(IBig::from(sig), e);
        let past_old_ceiling = |r: &Repr<10>| r.exponent() > isize::MAX / 2;

        crate::ziv_retries_reset();
        let sh = ctx.sinh_pi::<10>(&x, None).unwrap().value();
        assert!(crate::ziv_retries() <= 1, "sinh_pi @20 should certify within one retry");
        assert!(!sh.repr().is_infinite() && past_old_ceiling(sh.repr()));

        crate::ziv_retries_reset();
        let ch = ctx.cosh_pi::<10>(&x, None).unwrap().value();
        assert!(crate::ziv_retries() <= 1);
        assert!(!ch.repr().is_infinite() && past_old_ceiling(ch.repr()));

        // the pair shares one evaluation and certifies both parts together
        crate::ziv_retries_reset();
        let (sh2, ch2) = ctx.sinh_cosh_pi::<10>(&x, None);
        assert!(crate::ziv_retries() <= 1);
        assert_eq!(sh2.unwrap().value(), sh);
        assert_eq!(ch2.unwrap().value(), ch);
    }

    /// `acosh`/`atanh` compare their `±1` boundaries *exactly*: an argument that merely rounds
    /// onto the endpoint is not the endpoint, and both have a vertical tangent there — the true
    /// value is `√(2δ)` (acosh) / `ln(2/δ)/2` (atanh), not the endpoint's `0` / `∞`.
    ///
    /// The rows that *used* to be at risk here are covered: `acosh` takes `√((x−1)(x+1))` and
    /// `atanh` divides `2x/(1−x)` by an argument that rounds onto the pole, so both reach the
    /// retry path rather than certifying the endpoint — `acosh` via the loop's zero-candidate
    /// guard, `atanh` via `ln_1p_ball`'s pole guard (which previously panicked on the `±∞`
    /// ratio).
    #[test]
    fn test_hyperbolic_endpoints_compare_exactly() {
        for &p in &[20usize, 50, 100, 500] {
            // 10^-n is beyond the initial guard, so the argument rounds onto the endpoint at the
            // first work precision for p ≥ 100 (the case that used to snap); at p = 20/50 the rows
            // are plain correctness checks.
            let n = 3 * p / 2 + 5;
            let below = DBig::from_str(&format!("0.{}", "9".repeat(n))).unwrap(); // 1 − 10^-n
            let above = DBig::from_str(&format!("1.{}1", "0".repeat(n - 1))).unwrap(); // 1 + 10^-n
            let ctx = Context::<mode::HalfEven>::new(p);
            let hi = Context::<mode::HalfEven>::new(p + 60);

            let a = ctx.atanh::<10>(below.repr(), None).unwrap().value();
            assert!(a.repr().is_finite(), "atanh(1 - 10^-{n}) at p={p} is finite, not -inf");
            assert_eq!(
                a,
                hi.atanh::<10>(below.repr(), None)
                    .unwrap()
                    .value()
                    .with_precision(p)
                    .value(),
                "atanh(1 - 10^-{n}) at p={p}"
            );

            let c = ctx.acosh::<10>(above.repr(), None).unwrap().value();
            assert!(
                !c.repr().significand().is_zero(),
                "acosh(1 + 10^-{n}) at p={p} is not an exact 0"
            );
            assert_eq!(
                c,
                hi.acosh::<10>(above.repr(), None)
                    .unwrap()
                    .value()
                    .with_precision(p)
                    .value(),
                "acosh(1 + 10^-{n}) at p={p}"
            );

            // the exact endpoints keep their exact results
            let one = Repr::<10>::one();
            assert!(ctx
                .atanh::<10>(&one, None)
                .unwrap()
                .value()
                .repr()
                .is_infinite());
            assert!(ctx
                .atanh::<10>(&(-Repr::<10>::one()), None)
                .unwrap()
                .value()
                .repr()
                .is_infinite());
            assert!(ctx
                .acosh::<10>(&one, None)
                .unwrap()
                .value()
                .repr()
                .significand()
                .is_zero());
        }
    }
}