GPT-5.6 prompt caching broken, users paying cache-write fees with no credits

User reports that GPT-5.6 models show zero cached_tokens despite paying the 1.25x cache-write fee, reproducible with simple curl requests. The bug affects the sol/terra/luna model variants and has drawn 25 comments on Reddit.