Synthetic Kimi vocabulary: the shared fixture generates 256 byte tokens, then
adds merges.txt in rank order and the configured protocol markers.
prompts.json records token counts and SHA-256 of little-endian u32 token IDs
from SGLang's _encode_messages and moonshotai/Kimi-K3 revision
f831ab66814297da540d832a5235f8e904f29d06. Regenerate in a SGLang Python environment:
python tests/scripts/generate_kimi_parity.py