Rebuild benchmark binaries that compile client source directly

When a benchmark driver compiles the system-under-test's source files directly into its own binary, rather than linking a library that make rebuilds, running make after editing that source does not update the benchmark. A rerun then silently measures the old code, and results identical to before (byte counts within 1%) are the tell. Before trusting a before/after comparison, compare mtimes of the bench binary and the edited source, or make the runner call the bench's build step itself. Separately: a group-commit server that fsyncs a batch's distinct files can still lose to plain parallel per-request fsyncs on NVMe ext4, because the journal already merges concurrent fsyncs. Measure it before claiming the win.