fix(infra): capture a final runner metric sample at job end

What

Captures a final machine-metrics sample at the very end of a macOS runner job, so the charts reach the job’s end instead of stopping ~one interval short (the “Complete job” step currently has no corresponding data point).

Why

metrics-poll.sh loops sample → sleep 15s, and dispatch-poll.sh’s EXIT trap halts the VM (shutdown -h now) the instant ./run.sh returns — killing the sampler mid-sleep. So the last interval, including “Complete job” and the post steps, never gets sampled.

Confirmed on job 84735487285:

A ~13s tail (one interval) with no data.

How

• metrics-poll.sh: factor the per-iteration sample+POST into an emit_sample function and add a --once mode (one sample, no loop).
• dispatch-poll.sh: after ./run.sh exits, capture its rc, run metrics-poll.sh --once (network still up, before the VM halts), then exit with the runner’s rc.

Best-effort — it never affects the runner’s exit code, and a failed sample just leaves the tail as-is. The one-shot has no previous counters so its network deltas report 0, which is fine since end-of-job throughput is ~0.

Validation

• shellcheck clean on metrics-poll.sh; dispatch-poll.sh shows only its two pre-existing warnings (untouched). Both pass bash -n.

Scope

• macOS runner image only (this is where the observed gap is). The Linux kata sidecar has a similar, smaller tail on pod termination and can be handled separately.
• Takes effect for VMs booted from the next runner-image roll.

This is the follow-up to the now-merged runner-metrics work (#11595: axis alignment + 8h token TTL + clock sync).

🤖 Generated with Claude Code

No GitHub comments yet.