The container load planner packs cartons into a container one at a time. Each unit tries the free corners the packing has so far, lowest first, and takes the first one where it fits the walls, overlaps nothing, has enough support underneath and stays inside its stacking limits. Then the next unit does the same, and the whole pass repeats with different orderings to find a better load.
On a bench of 40 mixed loads, that test ran 6.4 million times and failed 5.9 million times. More than half of the failures, 3.56 million, could never have succeeded: the box overlapped something already placed, sat on top of too many layers, or would have wedged under a box above. Boxes are never taken out during a pass, so once a box of a given size can't go in a corner for one of those reasons, it never can. The planner was asking again for every later carton.
Now each corner keeps a small flag per box size. When a test fails for one of those permanent reasons, the flag goes up and later cartons of the same size skip that corner. Two failures stay unflagged on purpose: too little support underneath (a later box can fill the gap) and too much load on a box below (a new supporter changes how the weight is shared).
The random-load bench went from 4.2 s to 2.7 s, and three sets of 150 random loads ran 1.25 to 1.35 times faster. Every placement is the same: the plans hash to the same value before and after, and a new test pins 24 mixed loads to the old packer's output. A deliberately wrong version that also flags support failures breaks that test.
This matters on the page more than in CI: the planner runs in your browser when you click Plan, and mixed cargo with rotation and stacking limits is the case that got faster. A load of 1,600 identical cartons didn't change, because its time goes elsewhere; the perf note lists what to try next.