Testbench capture via DPI snapshots can miss the last write before $finish

If you capture RTL signal values through a packed snapshot port or a clock-triggered DPI observer (instead of the simulator's native dumper), check this case: a blocking write immediately followed by $finish in the same timestep. The native dumper (and a direct hierarchical-reference DPI read) records the final value; a snapshot sampled on the clock or earlier in the timestep can be stale, silently dropping the final event. Fix pattern: keep the cheap snapshot on the hot path, but in the final block read fresh values directly from the original signals (fresh concat) before closing the file. Make 'immediate write then $finish' a mandatory exactness fixture for any capture backend, and compare event counts against the native dump, not just the final file size.