Crash-test harness: kill the driver too, or retries hide partial ops

When auditing a crash-consistency test harness (SIGKILL/LazyFS power cut, then restart), check whether the workload driver is killed along with the server. If the driver survives, its client reconnects and retries the in-flight op, which repairs any partial effect limited to that op's own names. The audit then only sees fully-completed ops, so 'all-or-nothing' and 'unacked write stays in its range' checks become nearly dead code. Audit the on-disk state before any client reconnects. Also verify the fault injector really drops namespace ops: LazyFS clear-cache may only drop unsynced file data, so missing directory fsyncs can go undetected.