[codex] Add bitlift sidecar runtime and eval tools

[Chedrian07]

Summary

• add bitlift sidecar loading/binding for routed expert Q4 tensors
• add Metal-side sidecar routed dispatch buffers and route tracing support
• add GGUF/source conversion tools for base GGUF, HF FP4, and HF Base-FP8 source sidecars
• add local Korean/KMMLU/Think MAX/long-instruction evaluation tools and reports

Validation

• python3 -m py_compile tools/eval_ds4_project.py tools/eval_kmmlu_sample.py tools/bench_thinkmax_ds4.py tools/eval_long_instructions_v2.py tools/write_bitlift_sidecar_from_hf_fp8.py
• make -B ds4
• generated and loaded L10 Base-FP8 Q4 sidecars: Top64, Top128, Full256
• ran inspect and Think MAX route traces for all three sidecars
• ran structured Korean/control/exact-long evaluation
• ran KMMLU 300 sample
• ran Think MAX expanded30
• ran long instruction v2

Result Notes

The Base-FP8 L10 source sidecars are runtime-stable, but they are not the recommended quality winner. Current operating recommendation remains: base or LateStable5Q4 for general nothink chat, Layer10Q4 for Think MAX/KMMLU-style Korean experiments. The uploaded private HF artifact repository records the experimental sidecars and reports for reproducibility: https://huggingface.co/KCh3dRi4n/DeepSeek-V4-Flash-KR-L10-BaseFP8-Q4-Sidecar-GGUF