v0.16.3

Retention prunes only runs with a manifest. Up to 0.16.2, archives from before switchyard in a set's directory (or a run cut off halfway) were pruned like any other run once their date left the daily window; upgrade before that happens to yours. Also in this release (0.16.2): every dump waits for its database (2 minutes, 10 for scheduled runs), and so does a restore before it drops and imports one; bucket copies are tried again in every run; a failed backup run removes its own files; a deploy's pending-SQL check runs once more when it fails.