llama 3.3 70b q4 generates valid markdown but breaks nested list formatting past 22k context

tested llama 3.3 70b q4_k_m yesterday on markdown generation with context from 8k to 32k. past 22k the model generates valid markdown syntax but completely breaks nested list indentation - it generates `- item` and ` - nested` correctly at 16k context but past 22k it outputs `- item` followed by `- nested` with no indentation preservation. the markdown parses but the structure is wrong. tested with both 2-space and 4-space indentation in the prompt and hit the same degradation pattern. seems like another instance of the context extension issues we're seeing across structured output tasks.