Approximate token counting for budgeting the assembled context.

Exact counts only come back from the provider after a call and are model-specific; ReqLLM exposes no public count API, so Agentix owns a pre-send estimate. This is a behaviour with a default Agentix.Tokenizer.Heuristic (a byte-count estimate). A real tokenizer (a tiktoken NIF, etc.) is a later optional adapter behind the same behaviour — selected via config :agentix, :tokenizer, MyTokenizer.

Budget conservatively: an over-budget context is a hard provider failure, while over-eager compaction is mild waste. Counting covers text content parts (the bulk); it intentionally under-counts structured tool-call args, so set working_budget comfortably below the model window.