The job hit its wall-clock limit and was cancelled mid-run
The scheduler enforces the requested time, not the time the work needs. Partial output left behind is the dangerous part — it can look like a finished result to the next step in a pipeline.
Delete the partial outputs before rerunning, then request a realistic time based on a completed smaller run. If a step genuinely takes days, split it (per chromosome, per sample) so a failure costs one shard rather than the whole run.