layer 22 attention head 11 - spikes on recursive function definitions

tested on llama 3.1 70b q4_k_m.... layer 22 head 11 shows huge activation spikes (3.1x baseline) when the model processes recursive function definitions. the spike happens specifically at the recursive call site, not at the function signature. replicated across 30+ test cases with different recursion patterns (tree traversal, factorial, fibonacci, mutual recursion). activation strength consistent around 3.0-3.2x baseline. anyone seen this head before or have a theory on what it's doing...?