Yeah high density of traffic both reduces visibility and requires maneuvering among other vehicles. I've also noticed that FSD Beta's memory of arrow road markings that should have been visible even if just briefly isn't great when waiting in traffic at a red light sometimes causing it to think it must turn when it could have gone straight or vice versa.
As you suggest with potentially using signs that are more visible or improving the memory aspect could be possible solutions with varying difficulty to implement and potential benefit. This could be something Tesla has decided to defer as polish because while a wrong lane selection is annoying maybe requiring rerouting, it's not usually a safety issue.
As others have highlighted, a lot of poor lane selection is due to bad map data whether it's incomplete or wrong, and Tesla does seem well positioned to generate high quality lane count and turn lane data as FSD Beta is able to make those predictions on the fly but sometimes just a few seconds too late to be comfortable. So with this dynamic mapping approach, it could require rerouting the first time like an unfamiliar driver in a new area, and the experience could be more polished subsequent times. Combined with the rest of the fleet, it would be pretty rare for a "first time." One technical reason for something like this is Autopilot is already using map data even when it's bad, so having better map data doesn't require more compute from the FSD Computer to understand signs or have better memory.
Overall Tesla has been prioritizing safety and focusing on higher reach/impact changes with their limited resources (people and compute). It'll be interesting to see what Tesla focuses on when the incremental benefit to safety for some effort decreases to allow for more polish.