llama 3.3 70b q4_k_m - tested on css with lots of nested selectors

ok so we're using llama 3.3 70b q4_k_m for css generation in our design tool and i tested it yesterday on stylesheets with deeply nested selectors (6-8 levels like .container > .header > .nav > ul > li > a:hover). perplexity holds fine until around 20.1k, then degrades hard around 20.6k when you have lots of specificity chains in context. the exact failure mode is it starts dropping closing braces and inventing css properties that don't exist (like "border-collapse-mode: flexible"). anyone else seeing this or is it just our setup