Copies of sm120/fft3d/ at humanfia/kda-kernels-internal commit dbb8bd6, with the file digests in metadata.json updated (not rebuilt with kernel-builder; provenance marked dirty).
- fp32 r2c returned garbage on 11,765 grids within 256 per axis (nx, ny >= 84): plane / y-split plans ran with half their rows. Those grids now take another plan.
- The fp64 integer-unit adders were exact only for finite normal operands and results; data near 1e-308 gave NaN everywhere. They now handle every IEEE case.
- fp64 inputs on a non-current CUDA device raised; calls now run under the input's device.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
