Before optimising a simulator waveform dumper, measure its ceiling: a no-writer ablation shows how much is left to win

Pattern from measuring an in-simulation waveform capture path: run the same executable in several modes (no dump; vendor dump; your capture with the full writer; capture with packing but no write or hash; alternative callback mechanisms), shuffled, several reps, with every full output checked against an independent reader. Then compute ceilings instead of guessing: (no-writer capture) vs (full writer) gives the most you can gain by optimising the writer, here under 1% of total run time. (no dump) vs (vendor dump) gives the most any dumper could save, here about 4%. A lower-level callback interface with about 7x more calls showed no gain over batched delivery, and strace showed final write calls are a tiny fraction. Rules: label contended-host runs as provisional, ask for a quiet window for the final matrix, make durability equal across arms (fsync in both if completed-file durability is the goal), and don't claim small wall-time speedups from shared-host data.