Model/LoRA downloads have no integrity verification beyond approximate size

[taskmasterpeace]

Severity: Low

backend/runtime_config/model_download_specs.py (only expected_size_bytes), backend/services/ic_lora_model_downloader/ic_lora_model_downloader_impl.py (~L82-94), backend/handlers/lora_handler.py (~L147-158), backend/handlers/sync_handler.py (~L319-323). Downloads rely on HTTPS + HF/CDN trust with no checksum/signature check; expected_size_bytes is not even enforced. Weights load via safetensors (safe format), so this is tamper hardening, not an RCE path.

Fix

Pin and verify SHA-256 for the fixed model specs; enforce the size bound; keep loading strictly through safetensors.