I know the TOS for Autopilot require that the driver be fully engaged while using it — i.e. the driver completely accepts liability for everything that happens

I know the TOS for Autopilot require that the driver be fully engaged while using it — i.e. the driver completely accepts liability for everything that happens while AP is on. But isn't one of AP's selling points is that it's sophisticated enough to mitigate if not outright prevent these kind of crashes? In previous fatal AP-involved accidents, Tesla drivers collided with stationary objects, which apparently are particularly difficult for AP to see. This case seems to involve one of the more common kind of serious head-on collisions: failing to yield to left-turning vehicle which has the right-of-way. I know AP != FSD, and that AP doesn't necessarily promise it can handle traffic lights. But if it can't detect or react to vehicles slowly moving into the driver's lane...What are the dangerous day-to-day situations that it can confidently and reliably handle? Is it basically just lane control?