fix: read the logs once more before a goodput measurement completes

[dims]

What was wrong

handleSucceeded and handleFailed both build from the stored status via buildCumulativeFromStatus and never read logs again. Anything the workload wrote in the last sampling window was dropped — sampleInterval, 60 seconds by default.

Seen on hardware (test-plan P3.1): a script that logged up to iteration 100 was recorded as reaching currentStep: 90. The workaround at the time was to make the workload sleep before exiting, which is not something a user should have to know about.

It matters most on failure, where the last lines before a crash are the interesting ones.

The shape already exists

The bandwidth controller does exactly this. On JobSucceeded and JobFailed it calls handleRunning for one more parse before going terminal:

if cond := ...JobSucceeded...; cond.Status == metav1.ConditionTrue {
    if len(measurement.Status.Results) == 0 {
        if _, err := r.handleRunning(ctx, measurement, job); err != nil { ... }
    }
    return r.handleTerminal(...)
}

bandwidthmeasurement_controller.go:152-167. finalSample gives goodput the same shape.

The detail that makes it work

handleRunning throttles reads so a status update does not re-trigger one:

if last, ok := r.lastSample[key]; ok && time.Since(last) < interval {
    return ctrl.Result{RequeueAfter: interval - time.Since(last)}, nil
}

The final read lands moments after the previous sample, so it would normally be skipped — the throttle would silently defeat the fix. finalSample clears the recorded sample time first.

Two tests cover both halves, so the mechanism is pinned rather than assumed:

• TestHandleRunningIsThrottledWithoutFinalSample — a plain call is throttled and asks for a requeue
• TestFinalSampleBypassesTheThrottle — finalSample clears the sample time

Scope

Best effort, matching the bandwidth version: a failure in the final read logs and leaves the previous status in place, which is what would have happened anyway.

Verification

make lint 0 issues, make build clean, make test all packages pass. No golden file changes — the existing goodput cases read the same fixture logs, so one extra read produces the same result.

[dims]

/ok to test