The harness was lying
The report from play was simple: you get stuck in doorways, and sometimes you have to jump to get into a building.
So I wrote a test that walks a virtual player into every building on the map and checks whether they get inside. It reported 38 of 70 buildings unenterable. That is a catastrophic number, and it was almost entirely fiction.
After fixing the test, the real number is 7. Six of those have no door in the art at all. Most of that original number was the harness, not the game.
Four traps, each fatal on its own
Every one of these produced a confident, specific, wrong answer.
The floor had holes in it. The test tiled its ground from a slab prop, which spans a range that does not divide evenly into the 20m grid it was laid on. That leaves gaps. The walker fell into them and caught on slab edges, and the resulting failures had nothing whatsoever to do with the building being tested. It is one flat box now.
The walker did not move like a player. It drove a constant velocity straight into the collision resolver. The actual game accelerates toward a wish speed, applies gravity only when airborne, and checks whether you are on the ground before doing either. A test that moves differently from the game tests something that is not the game.
It walked into the centre of each wall. Doorways are rarely centred. The centre line of a wall is, more often than not, wall. Sweeping every 0.5m along all four faces instead took the failures from 37 to 15, meaning over half the remaining "unenterable" buildings had a perfectly good door that the test had simply never aimed at.
It required the walk to end inside. One shop has a 21.5m open front. You walk in, you keep going, you come out the other side, and the test counted that as blocked, because the walker finished outside. Checking the entire path rather than the final position: 15 down to 7.
What was actually broken
Fifty-six of fifty-six buildings with a real door can be walked into. I sabotage-checked this by filling a building solid and confirming the test notices.
Of the ten exclusions, eight have sealed art. One shop's widest clear span at body height, on any face, is 0.25m, verified against the actual model triangles. It is a facade, meant to be seen from outside. Collision that let a player inside it would be the bug.
Two are real faults I have not fixed: two buildings with genuine door-sized openings, 8.25m and 6.25m, that are still blocked. They are listed by name rather than hidden by loosening the assertion until it passes.
The lattice that looked like the answer
One promising fix was a finer collision lattice. It visibly worked on the worst building.
It worked for the wrong reason. A finer lattice thins the floor slab. At the coarse setting the floor is thick enough to block a step, and at the fine setting it is not. The player was squeezing through the floor, not walking through the door. Measured across all ten problem buildings, it fixed two, cost five to ten times more collision geometry, and left the other eight exactly as blocked as before.
Walk it, do not measure it
The lasting lesson is about a class of test rather than any one bug.
Every time I tried to answer "can a player fit through this?" by measuring the geometry, I got a wrong answer. A doorway-width probe reported one shop as passable when a body demonstrably cannot fit. The same probe separately reported a 5.07m opening that does not exist. It had found two deduplicated vertex coordinates far apart and read the space between them as a gap, with no triangle actually connecting them.
Geometry probes answer the question you encoded, which is never quite the question you asked. Walking a body through is slower and it is the only thing that has been right.