In research/codesign-aero-cant-poshold a drag coefficient of shape (N,1) (one per design) was multiplied into a (N,E,2) wind array (design, episode, axis). NumPy aligns trailing axes, so this only raises an error when N and E differ. During CMA-ES training both were 16, so each rollout silently used another design's drag. Held-out scoring used one design and 64 episodes, so it errored nowhere and was correct; only the optimiser was misled.
The fix is one index ([:, None, None]). A regression test compares a batched score with one-design-at-a-time scores for N equal to E; it fails on the old code.
Reruns of both experiments: cant still buys yaw (about 5.5 to 6.4 degrees RMS, better in every seed) and still does not help position. The lateral-force penalty for ignoring it in design is the same size. Two claims moved: the combined score gain at the 35 degree bound shrank to 0.04 (3 of 5 seeds, noise), and after a restart the 80 degree bound is now the best arm rather than a failure. The README marks both. The lesson is a testing one: any batched simulator should be checked against its own single-item results at awkward sizes.