[v0.2.0] Archive V100 comprehensive design-space results

[TruNcat3]

Summary

• formalize the operator-independent butterfly and FFT architecture spaces
• integrate optional cuFFTDx, TurboFFT, and VkFFT processing-unit paths
• archive V100 FFT exploration, NCU evidence, and the controlled 37-case comprehensive suite
• document current performance boundaries, reproducibility commands, and the next batch/length scan phase

Why

This freezes the current single-V100 research stage before beginning the next orthogonal batch and transform-length scan. It keeps architecture-level mapping separate from replaceable local processing units and preserves raw, summarized, and interpreted evidence together.

Results

• all 37 comprehensive cases pass correctness preflight
• FP32 FFT is at cuFFT parity for the measured logN 8 and 14 shapes, reaches 1.031x at logN 18, and reaches 0.892x at logN 20
• FP64 logN 16 remains at 0.642x cuFFT and is recorded as an open gap
• 36 cases are stable across their full trial range; one short cuFFT case is explicitly marked stable-with-outlier

Validation

• Release CUDA 11.8 / sm_70 build with cuFFTDx and VkFFT enabled
• CTest: 5/5 passed
• Python design-space and comprehensive tests: 32/32 passed
• repository checks: 29 Markdown, 8 JSON, and 25 Python files
• git diff --check