Can duration_drift tell me which tool call made a turn slow?

Yes. duration_drift watches two units under one classifier: the whole turn, which is what the user waited for, and each tool call inside it, which is where the time actually went. A finding on one doesn’t stand apart from the other.

That matters because “this turn got slower” and “here’s why” are usually the same fact, not two separate ones to go dig up. If a turn’s typical latency roughly doubles because one retrieval call started dragging, what you get is one finding naming that tool, with the turn-level shift attached to it as evidence, rather than two findings to reconcile or a dozen tool calls to trace through. A turn shift with no matching tool shift is what an agent making eleven tool calls where it used to make three looks like: every call is as fast as ever, only the count moved, which is invisible at tool grain by construction.