[agourakis82]

Complex acosh was built as (-imag(acos(z)), real(acos(z))), which leaves the lower half-plane on the wrong branch. It now returns the nonnegative imaginary magnitude of acos and the sign of imag(z). The same change keeps the signed real-axis asin correction, so acosh on the cut agrees with std::acosh. An 8 by 8 grid, including (2, -1) and both sides of the cut, is compared with std::acosh with and without fast-math.

Part of #1432.

Written by me with AI assistance.

[DiamonDinoia]

This is okay but still I need to see a rebase.

[agourakis82]

Published in f1ad98f.

I made this candidate self-contained by including the necessary real-axis asin correction already implemented in #1437. Complex acosh uses acos, which in turn uses asin, so the inherited real-axis asin result must preserve both its nonzero imaginary magnitude and the relevant signs. This is the existing #1437 implementation correction reused here.

The existing typed acosh fixtures now include imaginary parts +0 and -0 alongside +1 and -1, over the same eight real values. Numerical comparisons execute in strict and fast-math. Explicit imaginary-zero sign checks remain in strict mode.

On macOS ARM64 with Apple Clang 21 and C++23, the scoped complex hyperbolic and complex trigonometric tests pass four cases in each mode, with 96 assertions in strict mode and 64 in fast-math. Restoring the inherited pre-fix asin implementation in a receipt-local header copy makes the updated regression fail in both modes. A separate C++23 comparison of acosh(±2, ±0) against std::acosh passes all eight float/double cases in both modes for this candidate.

Prepared with OpenAI Codex assistance.

[DiamonDinoia]

Superseded by #1456