qwen 2.5 32b generates working regex but catastrophic backtracking past 18k context

tested qwen 2.5 32b q4 yesterday at 18.3k context with email validation regex generation.... the pattern looks syntactically correct but has catastrophic backtracking on edge cases like nested quotes or really long local parts at 15k context same prompt generates efficient patterns with proper anchoring. past 18k the patterns work on basic test cases but hang on adversarial inputs is this context degradation or is qwen just bad at regex complexity analysis