//! Kernel implementations for quantize, dequantize, cosine, and residual
//! helpers (Phase 20).
//!
//! The `scalar` submodule is the canonical reference and is always
//! compiled. Under `feature = "simd"`, architecture-specific submodules
//! (`avx2`, `neon`) provide wrappers that delegate to scalar until full
//! intrinsic paths land. The `portable` submodule exists as a documented
//! fallback — on MSRV 1.81 `core::simd` is unavailable, so it also
//! delegates to scalar.