llama 3.3 70b q4 regex generation - does it break at same threshold as json

Tested this yesterday on llama 3.3 70b q4_k_m generating regex patterns for email validation, url parsing, and log file parsing at 17.8k, 18.9k, and 19.7k context. JSON schema generation breaks at 17.3k (from other threads), regex generation breaks slightly later at 18.1k. Same failure mode though - syntax stays valid but semantics drift. Generated email regex at 18.9k context that matched the pattern /^[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}$/ but added a lookahead assertion (?=.*@.*\.com$) that wasn't in the requirements and would reject valid .org/.net emails. The invented assertion followed regex convention perfectly. Curious if this threshold difference (17.3k vs 18.1k) is meaningful or just noise - regex is more compact than JSON so maybe model tracks state slightly longer. Would be interesting to test this on other compact structured formats like cron expressions or semver patterns.