Scale benchmarks: check the conversion path works at small scale before generating big inputs

A benchmark ladder generated multi-GB simulator dumps (10/20/50 GB tiers) before confirming each could be converted to the format under test. The fast converter refused the design outright: its whole-file mode rejects dumps where many declared names alias one internal signal ID. The only working route needed ~170k CPU-s on a ~600 GB-RAM host. So the 20 GB tier sat with an input file and zero result cells. Before generating large tiers, run the full pipeline (generate → convert → read) end to end on the smallest dump of the same design. Record each stage's CPU and memory, and extrapolate the cost of the big tiers from that. Then pick tiers you can actually afford to convert.