docs: full BF16 rewrite spec + plan — zero F32 on GPU
Spec: every CUDA kernel, every GPU buffer, every cuBLAS call → BF16.
37 HOT kernels + 11 WARM + 7 COLD = 55 total kernel files.
12 implementation tasks across 4 phases.
Phase 1: Foundation (buffer types, weight sets)
Phase 2: cuBLAS GemmEx BF16×BF16→BF16 (forward + backward)
Phase 3: Optimizer + loss kernels (Adam, C51, MSE, spectral norm)
Phase 4: Environment + auxiliary (experience, backtest, IQN, attention)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
136696d8b8