[zhangzhenghao]

Summary

• Add native C++17 Hugging Face GGUF downloads with clap download -hf owner/repo[:quant] and the same resolver in clap-model.
• Resolve immutable commits and paginated repository trees; support exact filenames, quant selection, HF cache layout/environment variables, bearer tokens, and offline reuse.
• Stream/resume downloads with validated ranges and sizes, bounded retries, LFS SHA256 verification, and credential-safe redirects.
• Validate cache paths across platforms and recover interrupted snapshot publication; fall back to hard links when symlinks are unavailable.
• Keep existing dummy embedding/server interfaces unchanged; reject split GGUFs explicitly because the model loader does not support them.

Validation

• Full local CTest: 18/18 passed.
• HF HTTP integration suite: 23/23 passed against both static and shared builds, including downloaded-model inference and offline equivalence.
• Shared build with BUILD_TESTING=OFF and staged install verified (existing shared install requires runtime library-path configuration).
• Independent code review passed; git diff --check passed.
• Native Windows/macOS validation is left to the CI matrix.

Notes

• HTTPS requires OpenSSL development libraries at build time.
• Downloads consume existing GGUF files, not safetensors/PyTorch conversion.
• An abandoned repository download lock requires manual cleanup as documented.