Offline DeepSeek input counting resources

Source repository: https://github.com/deepseek-ai/deepseek-recipe
Pinned revision: 8cadfede7063c896b944e7bae05daa3549ae97ea
Tokenizer profiles: v4 and v41; exact byte hashes are in manifest.json.
Official renderer sources are retained as provenance data and are never executed.
The local renderer supports only one plain-text system message plus one user message,
without tools, images, history or remote fetching. Counts include JSON output format,
reasoning-effort instructions, BOS/role tokens and the generation prefix.

DeepSeek recipe code and tokenizer resources: MIT; retain both supplied licenses.
Native tokenizers runtime: https://github.com/huggingface/tokenizers
Version 0.23.2; runtime wheel and all declared dependencies are hash-locked in
requirements-tokenizer.lock. Native/tokenizer dependency licenses belong to those
distributions and must be retained by the packager.

The count is exact for the pinned local template/tokenizer and is not an observed
remote API usage value. A different model/template/runtime remains unsupported.
No customer files, prompts, credentials, model weights or API responses are included.
