I Tested GPT-OSS-20B Against the Field. The Architecture Explains Everything

2 days debugging a 20B LLM on AWS. CUDA mismatch, OOM on g4dn, slow cold starts. Here is what fixed it.