write_scale never measures large repositories. The records it seeds go straight into the store and never reach the built repository: after 1,000 seeded puts, an export of the repository is 1.6 KB. So the benchmark's "records per repository" axis measures small repositories at every setting. Any cost claimed from it for large ones isn't backed by it. That includes the build costs quoted in the max_records_per_account docs, if they came from this benchmark.
Found at 73ddde8 while fixing the benchmark's account-creation ceiling (https://tangled.org/permadeath.com/didbot/pulls/1219).
Fix: seed through the same write path a client uses, or build the repository from the seeded records before the timed phase. Then assert the export size grows with the seeded count, so the benchmark fails if seeding stops reaching the repository again.
Fixed by https://tangled.org/permadeath.com/didbot/pulls/1253, now on main.