feat: add --duration flag for self-terminated load runs

[bdchatham]

Summary

Add --duration flag so seiload self-terminates cleanly inside K8s Job activeDeadlineSeconds, producing pod exit 0 → Job condition Complete instead of K8s-mandated Failed/DeadlineExceeded.

Why this exists

Followup to #30. The exit-code fix in #30 is correct in isolation (graceful SIGTERM → exit 0), but on Kubernetes Jobs with Job-level activeDeadlineSeconds, the K8s Job controller sets condition=Failed, reason=DeadlineExceeded regardless of the container's exit code:

Once a Job reaches activeDeadlineSeconds, all of its running Pods are terminated and the Job status will become type: Failed with reason: DeadlineExceeded.
— Kubernetes Job docs

A run on harbor's nightly namespace today (workflow 25189102396) confirmed this: seiload's pod was SIGTERMed at the deadline and likely exited 0 (post-#30), but the Job condition was set:

"reason": "DeadlineExceeded",
"message": "Job was active longer than specified deadline"

This means kube_job_failed=1 and KubeJobFailed keeps firing — the original symptom that motivated this whole investigation. The actual fix has to make seiload self-terminate before K8s decides "DeadlineExceeded."

What changes

rootCmd.Flags().Duration("duration", 0, "Run duration; the load test ctx is canceled after this elapses, the existing graceful-shutdown path runs, and the process exits 0. 0 means run until SIGTERM/SIGINT.")

if duration, _ := cmd.Flags().GetDuration("duration"); duration > 0 {
    log.Printf("⏰ Run duration: %s", duration)
    var cancel context.CancelFunc
    ctx, cancel = context.WithTimeout(ctx, duration)
    defer cancel()
}

When --duration is set:

1. The load-test context is wrapped with WithTimeout.
2. After duration elapses, ctx is canceled with context.DeadlineExceeded.
3. Background tasks (dispatcher, logger, block_collector) unwind via ctx.Done().
4. service.Run returns the wrapped DeadlineExceeded error.
5. Final stats emit, EmitRunSummary runs, post-summary flush sleeps for 45s.
6. Existing boundary check from fix: exit 0 on graceful SIGTERM/deadline shutdown #30 (errors.Is(err, context.DeadlineExceeded)) clears the error.
7. Process exits 0.

The existing post-summary flush delay still runs by design — it sits after service.Run returns, in the cleanup pipeline.

What doesn't change

• Default is 0 (unlimited) so existing callers without the flag are unaffected.
• SIGTERM/SIGINT handling in main.go:347-352 is untouched and still works.
• Exit-code semantics from fix: exit 0 on graceful SIGTERM/deadline shutdown #30 (Canceled OR DeadlineExceeded → exit 0) already cover both internal-timeout and external-SIGTERM paths.

Test plan

• GOWORK=off go build . passes
• GOWORK=off go vet ./... clean
• Companion platform PR will pass --duration=${DURATION_MINUTES}m to seiload args; tomorrow's nightly will verify Job condition flips to Complete

Companion change

Will follow with a small PR on sei-protocol/platform that:

1. Bumps the seiload image to the new SHA after this merges.
2. Adds --duration=${DURATION_MINUTES}m to seiload args in clusters/harbor/nightly/templates/seiload-job.yaml.
3. Optionally raises JOB_DEADLINE_SECONDS slightly to keep activeDeadlineSeconds as a backstop only.

🤖 Generated with Claude Code