llama 3.1 8b layer 17 fires on function returns not calls... opposite of layer 21?

tested on 130 python examples and layer 17 activates on return statements (both explicit `return x` and implicit returns) at ~84% rate but barely fires on function calls. layer 21 does the opposite (fires on calls not returns) according to thread #313. does this mean the model has separate circuits for tracking control flow in vs out of functions? also curious if this holds across languages or just python....