Written by an AI agent working in this repository.
Two sessions on 2026-11-18 reported the same local failure in research/sensing-selfcal: test_warm_start_matches_cold_on_a_converged_set asserted 42 <= 22 and failed. No CI workflow runs that project's tests, so the failure only ever showed up in somebody's container.
What the test claimed
Milestone 4k asked whether starting the camera self-calibration from a fleet's nominal focal length (527.1 px) lands where milestone 4j's cold start (1.3 times the Zhang estimate) does. The test checked one well-conditioned 6-view set on TUM's fr3 camera and made two claims. The first is that the warm start lands on the same f to 1e-6. The second is that the warm start takes no more function evaluations than the cold one did. For the cold count, it read the evaluation count stored in milestone 4j's results.json.gz.
What reran
The corners are read from stored JSON, so the inputs were the same bytes. I reran the solver for that set and the first 40 other 6-view subsets on SciPy 1.18.1 / NumPy 2.5.3:
| stored | rerun here | |
|---|---|---|
| test set, cold start | 22 | 47 |
| test set, warm start | 12 | 42 |
| 41 sets, cold count equal to stored | 22 of 41 | |
| 41 sets, warm count equal to stored | 16 of 41 | |
| 41 sets, warm no more than cold (same run) | 35 of 41 |
Levenberg-Marquardt through MINPACK takes a different path on a different SciPy or BLAS build, and stops after a different number of steps at the same minimum. Every f still matched its stored value to 1e-6. The milestone's findings rest on f, the far solves and the false-alarm rates, not on evaluation counts, so none of them move.
The fix
The test now runs the cold solve itself, in the same process, and compares the two counts there: 42 against 47. Against the stored file it checks only f, which does reproduce. The full project suite passes (146 tests, about two minutes). The rule generalises: compare an iteration count only with one counted in the same environment, and pin only stored results that rerun to the same value.