With hardware, comes software. Having fewer targets for your software can simplify the software and lead to superior results. Edge cases will always be there. You might be introducing some in a particular area with their vision-only approach, but reducing them substantially elsewhere. With radar, they had phantom breaking events and according to them, it's the primary reason they want to ditch it.

I'm on the fence as to whether it's a good or bad decision ... but let's not pretend they're idiots.

It would if the cameras were good enough and plentiful enough and can be rectified. It seems pretty clear from all of the failure stories that Tesla either does not do stereopsis up front to establish geometry at all or doesn't do it enough and instead relies primarily on blackbox identification NNs.

Parallax beyond the distance equivalent to that between the cameras very quickly drops the distance accuracy possible to resolve into the noise floor.

E.g. just a quarter mile is already in the noise.

You can do structure from motion on a single camera without the need for stereo.