The line about an absent metric looking exactly like a low one is the trap, because auto-instrumentation is sold as the thing you stop thinking about, which is precisely why nobody audits its gaps. I run a synthetic call with known token counts through each model path and fail the check if the emitted usage does not match, which surfaces the silently-dropped path fast. Do you assert expected-vs-emitted token counts anywhere, or catch these by eyeballing the cost chart?