[codex] Add v0.5 numeric-regime mapping study

[TruNcat3]

Summary

• add public FP16/BF16 storage and native-width/FP32 accumulation contracts for FFT, FWHT, and Structured 2x2, including V100 BF16 emulation labeling
• add the fixed-V100 numeric-regime generator, quick screening, full boundary follow-ups, paired NCU attribution, adaptive coverage, and an abstaining piecewise selector
• add the representative library/base/search matrix for FP32 FFT/FWHT and 60-bit NTT, with checked-in raw data, reducers, tests, and layered documentation

Evidence

• quick screen: 736 cases, 2,208 samples, 29 confirmed and 67 ambiguous crossovers
• full follow-up: 23 same direction, 3 reversed, 3 unconfirmed; all 9 length crossovers reproduce
• adaptive coverage: 99 anchors, 66 stable, 33 near-ties, 0 reversals
• piecewise selector: 27.74% coverage, 98.84% top-1, 1.00013x geomean regret, 1.01112x worst regret in the selected region
• ten-shape matrix: searched/base 2.349x and searched/library 1.109x geometric mean; FFT 1.015x, FWHT 1.054x, NTT 1.313x against matched libraries

Validation

• python3 -m unittest discover -s tests -p 'test_*.py': 99 passed
• /usr/bin/ctest --test-dir build-11.8 --output-on-failure: 47 passed
• python3 scripts/check_repository.py: passed
• git diff --check: passed
• rebuilt the numeric NCU analysis, final piecewise selector, and three-way tables from checked-in raw records

Evidence Boundaries

• V100 with CUDA 11.8 is the only measured GPU target; cross-GPU transfer remains future work
• the piecewise selector abstains outside calibrated stable intervals and global runtime status remains measurement-required
• GPU-NTT rows in the three-way matrix are archived same-V100 matching-protocol measurements, not interleaved with the current FFT/FWHT refresh
• the representative matrix does not claim universal superiority over every library shape or semantic contract