[agourakis82]

atan2 forms an invalid ratio when both inputs are zero or both are infinite, producing NaN angles instead of the signed zero/pi or quadrant angle required for those pairs. The zero case also gives log(0 +/- 0i) a NaN imaginary component.

Use safe temporary division operands for these pairs. Preserve the original signs for zero-pair selection and let the existing quadrant calculation handle the unit ratio for infinite pairs. Add mixed-lane float/double coverage for zero pairs, infinite pairs, axes, ordinary quadrants, single-infinite controls, NaN propagation and complex log at signed zeros. Nonzero angles are compared numerically; zero values and their signs are checked exactly.

Validation (C++23):

• ARM64: the additional infinity regression fails against preceding PR commit 4bc84b2 (60 assertions); the expanded affected units pass with the final patch (10 cases / 732 assertions). ASan/UBSan passes the same units.
• Native AVX2/FMA, GCC13.3 on Intel Xeon 6730P: identical expanded tests fail against upstream master (312 assertions) and preceding PR commit 4bc84b2 (120 assertions), then pass with the final patch (10 cases / 1464 assertions). CPU-only Slurm20310 completed with exit0:0.
• Floating-environment witnesses cover all four uniform zero pairs, all four uniform infinite pairs, and mixed exceptional/ordinary lanes. Final ARM64/native-AVX2 witnesses report no FE_INVALID or FE_DIVBYZERO; the preceding implementation returns NaN and raises flags for infinite pairs. No general exception-freedom claim for other inputs.
• Native AVX2 composition with Preserve the imaginary zero sign in complex log1p #1438 passes 12 cases / 2112 assertions, including both endpoint zero signs in log1p(-1 +/- 0i). This PR remains independently based on master; Preserve the imaginary zero sign in complex log1p #1438 preserves the imaginary sign during log1p translation.
• Earlier zero-only validation at 4bc84b2 also passed ARM64, ASan/UBSan, translated SSE2 and native AVX2. Current expanded counts above refer to the additional infinite-pair coverage.
• Formatting and whitespace checks pass.

Related to the endpoint found while investigating #1432. Exceptional-pair handling is excluded under fast-math; infinity coverage is guarded when infinities/NaNs are disabled. Broader overflow, underflow, nonfinite accuracy, a full-suite pass and all-architecture validation remain outside this patch.