ResourceHasRestartAnnotation only supports Deployments (and the step text is deployment-specific). Renaming this to DeploymentHasRestartAnnotation (and updating the step binding) would make the intent clearer and avoid implying it works for arbitrary resources.

This comment hard-codes an assumed reconciliation period (“controller runs every 10s”). If reconciliation is event-driven (e.g., triggered by the rollout restart changing the Deployment), this note is misleading and the fixed 30s sleep may be unnecessary. Consider rewording to avoid a specific cadence and/or replacing the fixed sleep with waiting for an observable reconciliation signal.

The tag @~BoxcutterRuntime won’t be interpreted as “skip when BoxcutterRuntime is enabled”. CheckFeatureTags only skips when a scenario tag matches a known feature gate name and that gate is disabled; ~BoxcutterRuntime isn’t a known gate, so this scenario will still run under BoxcutterRuntime and may fail (per the TODO). Consider either removing this tag and relying on the test runner’s tag filters, or extending CheckFeatureTags to treat tags prefixed with ~ as “skip when the corresponding feature gate is enabled”.

On kubectl failure, this error drops useful diagnostics (stderr). Other steps include stderrOutput(err) in failures, which helps a lot when debugging CI flakes. Consider appending stderr (and possibly the command output) to this error message.

k8sClient output may contain trailing newlines/whitespace; checking out == "" can be brittle. Elsewhere in this file empty kubectl output is checked with strings.TrimSpace(out) == "". Consider trimming before the emptiness check to avoid false negatives/positives.

This step/function is deployment-specific, but the name ResourceHasRestartAnnotation reads like a generic resource assertion. Renaming it to something deployment-specific (e.g., DeploymentHasRestartAnnotation) would make failures and step registration clearer.

The comments here assume a fixed reconciliation cadence (“OLM actually runs every 10 seconds”), but the controllers appear largely event-driven and don’t generally requeue on a fixed interval. This can mislead future readers and encourages fixed sleeps. Consider rewording to avoid stating a fixed period, and (if possible) prefer waiting on an observable reconciliation signal rather than a fixed delay.