Split a long text into token-sized chunks for RAG ingestion. Chunks are measured in REAL tokens (tiktoken o200k_base/cl100k_base), honour a token overlap, and cut on sentence, paragraph or word boundaries (strategy). Returns each chunk text, its exact token count and char offsets. The natural composite of token-count; counts are OpenAI-style estimates (no Claude tokenizer exists). Provided text only; nothing is fetched. Over 1 MB is truncated.