Summary

• Add new model configs: GPT 5.4 (pro/base/mini/nano), Grok 4.20, Qwen 3.5, DeepSeek v3.2, GLM-5, and older baselines (gpt-4o, gpt-4, gpt-3.5-turbo, claude-3-haiku)
• Add make score, make upload, and make run-benchmark Makefile targets with score_table.py CLI
• Fix resume logic to requeue failed/in-progress tasks that have no submission file on disk
• Extract _prepare_responses_kwargs() in OpenAI base adapter: default store=False, map reasoning_effort config to Responses API reasoning.effort, fix response parsing to skip reasoning items
• Filter HF uploads to JSON only to prevent checkpoint files from getting uploaded (allow_patterns=["*.json"])