Implements a generic temporal smoothing calculator using the OneEuroFilter algorithm. The header provides a reusable template function `SmoothLandmarks` that applies low-pass filtering with adaptive cutoff frequency based on velocity. This is a non-trivial signal processing pattern that can be extracted and applied to any time-series landmark data (e.g., hand, face, pose) with minimal adaptation. The implementation is self-contained, uses only standard types and the OneEuroFilter utility, and demonstrates clean separation of filtering logic from calculator boilerplate.

Provides a type-safe, variadic template for merging multiple input streams into a single output stream. The `Merge` function uses compile-time tag dispatching to handle arbitrary numbers of inputs, producing a vector of packets. This is a reusable building block for graph construction that avoids runtime overhead and demonstrates advanced C++ template techniques (parameter packs, constexpr evaluation). It can be extracted and used in any MediaPipe-like pipeline framework.

A concrete calculator that applies the OneEuroFilter to smooth landmark positions over time. It uses a velocity-based adaptive cutoff to reduce jitter while preserving low-latency response. The implementation is self-contained, with clear separation of filter state per landmark, and includes configurable options for min cutoff, beta, and derivate cutoff. This is a production-quality implementation of a common computer vision problem that can be reused in any landmark tracking pipeline.

Defines the `Tensor` class, a unified abstraction for multi-dimensional arrays that supports CPU, GPU (OpenGL, Metal, AHWB), and WebGPU backends. It provides zero-copy views, explicit memory management (pools, shared file descriptors), and type-safe accessors. The design is non-trivial: it uses a pimpl pattern to hide backend-specific implementations, supports on-device tensor operations, and integrates with MediaPipe's memory manager service. This component is highly reusable for any ML inference framework needing cross-platform tensor handling.

Introduces the `Node` class template for defining calculators with compile-time contracts. It uses variadic templates and constexpr strings to specify input/output ports and options, enabling static validation of graph connections. This is a significant improvement over the runtime protobuf-based contract system, demonstrating advanced C++ metaprogramming (type lists, tag dispatching, constexpr evaluation). The pattern is reusable for any framework that needs type-safe, declarative node definitions.