refactor: heel_f64x8 uses crate::simd::F64x8 polyfill, add SIMD cosine Replace raw std::arch intrinsics with crate::simd::F64x8 polyfill. Automatic dispatch: AVX-512 (native __m512d) → AVX2 (2×__m256d) → scalar. Consumer writes crate::simd::F64x8 — polyfill handles tier selection. Added SIMD cosine kernels using F64x8 FMA: cosine_f64_simd() — single-pass dot + norm_a + norm_b via F64x8 cosine_f32_to_f64_simd() — f32 input, f64 precision cosine dot_f64_simd() — F64x8 FMA dot product on f64 slices sum_sq_f64_simd() — F64x8 sum of squares 12 tests passing (6 HEEL + 6 cosine). https://claude.ai/code/session_01ChLvBfpJS8dQhHxRD4pYNp#75