Make a side-effecting tool safe to replay

eve's docs give two reasons this tool can run twice for one intended action. First, after a crash an interrupted step runs again from the checkpoint before it, and a call whose start hadn't reached the stream yet runs again with it. Second, a call that eve reports as interrupted may already have taken effect, and the model can then call the tool again, which arrives as a fresh call with a new callId. A world that redelivers a step while the earlier attempt is still running can even run both at once. So the protection has to live in the write itself: give each intended action a stable operation ID and record it before or with the effect. The decision that matters most is what that ID is made of. callId alone fails the second case, so the key is the matter plus the normalised enquiry text.