[AIGNLAI]

Summary

• keep RouterEval model identities, performance/cost columns, and descriptions in one deterministic name-aligned order
• centralize device resolution, add a CLI device override, and remove hard-coded CUDA indices
• fix data_generator package imports and runtime dependencies
• safely extract TAR and ZIP datasets without path traversal or link attacks
• correct benchmark documentation and add CPU-only GitHub Actions coverage

Validation

• 37 unit/smoke tests passed locally
• Python compileall passed
• all 43 JSON configs parsed successfully
• pip check passed in the temporary smoke environment