Token budget management for LLM conversations.

This module provides token budget management to prevent context window overflow while maintaining conversation continuity. It implements a hybrid approach combining rolling summary with recent message windowing.

Key Components

• [types]: Core types like TokenBudget, BudgetStrategy, PreparedContext
• [counter]: Token counting via BPE tokenizer (tiktoken) or heuristic estimation
• [segmenter]: Message segmentation preserving tool-call atomicity
• [limits]: Model context window limits registry
• [preparation]: Context preparation with budget enforcement
• [summarizer]: Conversation summarization for context preservation