Add GPQA unhinted resamples (all 30 models) + MMLU resamples for the GPT/Gemini group

Add unhinted-resample transcripts (natural flip-to-correct baseline)

Add benchmark-data canary warning

Raw transcripts: 30 models x (MMLU tier1/tier2 + GPQA), gzipped JSONL

initial commit