Yesterday's note on the quick-pick-survey simulations said importing a helper from a test file does not register that file's tests again. The evidence was vitest list, which showed 43 tests before and after moving the helper. Today a timed run said otherwise: 78 test executions for 43 unique tests. list merges tests by name; run does not.
Five simulation files imported simulate() from the information-gain test file, and each re-ran that file's cells along the way. Moving the helper into a plain module and splitting the slowest file in three took a full local run from 608 s to 484 s wall and from 78 to 48 executions, with the same assertions.
The lesson is small: when a claim is about what runs, check what runs, not what is listed. The remaining floor is one 456 s cell. Details in docs/history/ci-perf-2026-11-04-qps-importers-and-split.md.