Fabric hop latency: baseline before BI-3

When to read this: You are doing BI-8 (ADR-0020’s kill-criteria check) and need the number the new path’s hop latency is compared against.

Recorded by AGE-632 (BI-0) on main at e432477, before BI-3, with the swarm test kit (docs/swarm-test-kit.md).

What a hop is

From the leader’s invoke_agent call to the broker handing the child its task frame, both read off tracing timestamps in the leader’s process: the kit logs swarm kit: invoke_agent call just before the call, and LocalRunner::run_task logs Delegated a task to a local worker right after submit_task writes the frame. On this path a hop spawns a chatty-tui worker process and waits for it to register, so process start-up dominates.

200 scripted delegations, one at a time, each answered by one Text reply from the fake model.

cargo test -p chatty-tui fabric_hop_latency_baseline -- --ignored --nocapture

Result

Debug build, Linux, 4 vCPUs (the cloud container this issue ran in).

Baseline: p50 463.7 ms, p95 523.8 ms (run 1). BI-8’s metric is new p95 − baseline p95, measured with measure_hop_latency unchanged on comparable hardware; run 2 shows the run-to-run spread (≈ 10 ms at p50 and p95).