The crash was in the search
This crash had been closed once already. On 6 September it was traced to the mod asking the physics engine questions from the thread scripts run on, far too many of them, around twelve hundred per call, and it went away when the questions were cut down. Thirty calls without a crash. Then, at the end of September, it came back, at the same address inside the physics engine.
On 3 October it came back four times: 13:29, 16:05, 16:08 and 16:17. The first theory was a unit on its way to the player, because there was one in the log near the first crashes. The 16:17 crash killed that theory: the car had arrived well before.
What all four had in common had been sitting in the log the whole time. Every one landed in the exact second dispatch went looking for a place for the next call, a search it had scheduled fifteen seconds earlier. That fifteen was itself a test: that day, calls had been sped up from one a minute to one every fifteen seconds, which means four times as many searches. The working theory: every search carries a small chance of a crash, and four times the searches make it four times as likely.
The fix was not to make the search careful. It was to make it ask physics nothing at all. It no longer casts a ray — an invisible line — to check for a roof overhead, and it no longer probes the street. It looks at the traffic instead: a car driving 150 to 200 metres away, or a parked one, or walkable ground with open sky if there is no traffic at all.
Then it was left running on purpose, still at one search every fifteen seconds, for about an hour of driving around the city without accepting a single call. Ninety-two searches in a row, no crash. Before, it was one crash in roughly every seven to ten searches.
What is not proven is which question was the guilty one. The roof ray is the obvious suspect, and only a suspect: nobody measured the instant of a crash, because the crash takes the end of the log down with it. The fix works. The theory behind it is still a theory, and it stays written that way.