Prepare v0.3.0 counter-calibrated mapping artifact

[TruNcat3]

Summary

• formalize cuButterfly as a two-dimensional, four-factor space-time mapping model with logical decomposition separated from physical execution groups
• add generated FFT processing units, vectorized and FP64 online paths, counter-calibrated selectors, matching external baselines, and reproducible V100 analyses
• rewrite the root and architecture documentation around design motivation, GPU mapping semantics, evidence boundaries, and an abstract mapping diagram

Current evidence

• selector: 93.88% top-1, 100% top-3, 1.0049x geometric-mean regret over 49 internal shapes
• FP32 FFT: 1.077x cuFFT at logN=18 batch 16; saturated logN=20 remains 0.962x/0.954x at batch 8/16
• FP64 FFT: higher median throughput on 20/25 stable logN=14..18 length/batch shapes; all 13 stable logN=17/18 shapes are faster
• FP64 NCU: strength-reduced XOR addressing cuts prefix warp instructions 26.8% and prefix replay 8.3% versus the prior XOR kernel
• cross-GPU transfer remains explicitly deferred and is not claimed

Verification

• ./scripts/reproduce_v100_analysis.sh
• python3 scripts/check_repository.py
• cmake --build build -j
• /usr/bin/ctest --test-dir build --output-on-failure (39/39 passed)

Release note

This is a draft release PR. Version metadata and the GitHub v0.3.0 tag/release should be finalized only after review and merge; the branch does not commit directly to main.