Real BPE token counts, natively in Elixir — no Python/tiktoken venv.

Uses a vendored cl100k_base-equivalent tokenizer (priv/tokenizer/cl100k_base.tokenizer.json, a conversion published at the Xenova/gpt-3.5-turbo HF Hub repo) via the tokenizers Hex package (Rust-NIF bindings to HuggingFace's tokenizers crate — already resolved in this project's dependency tree via bumblebee, added here as an explicit, non-optional dependency since the benchmark tool needs it independent of the optional ML stack). Cross-checked against real tiktoken's cl100k_base encoding on sample text: identical counts.

Loaded once per BEAM instance via :persistent_term — no network access at benchmark-run time, and no per-call reparse of the ~4MB vocab file.