A self-contained UTF-8 streaming accumulator that buffers partial codepoints from token pieces and emits only complete characters. Handles multi-byte sequences split across arbitrary boundaries, with flush() for trailing fragments and clear() for reset. This is a reusable utility for any Dart project that needs to decode streaming UTF-8 bytes (e.g., from LLM tokenizers, network streams, or file readers) without corrupting multi-byte characters.

Defines a sealed class hierarchy for generation stream events: TokenEvent (sampled token with id, bytes, text, position), ShiftEvent (context shift notification), and DoneEvent (terminal with stop reason, counts, trailing text). This pattern cleanly models a streaming generation loop where consumers can pattern-match on event types. The design is reusable for any streaming token generation system (e.g., chat bots, code completion) and demonstrates idiomatic Dart sealed class usage.

A binary codec for serializing and deserializing LLM session state (KV cache, token history, metadata). Uses magic headers, versioning, FNV-1a checksums, and JSON metadata to ensure compatibility on reload. The encode/decode functions are self-contained and could be extracted for any project needing a robust, self-describing binary state format with validation. Demonstrates careful design for forward/backward compatibility and error discrimination.

Implements synchronous speculative decoding over two LlamaContexts (target and draft) sharing a vocabulary. Supports both greedy (temperature=0) and exact stochastic speculative sampling with residual resampling. The algorithm is well-documented and includes proper KV cache rollback, acceptance rate tracking, and reproducibility via seed. This is a complex, non-trivial implementation that could be extracted as a standalone speculative decoding module for other LLM inference engines.