+1 on this.
I thought I was going crazy.
This issue has been added to the Audition bug database.
Spoke with the main developer on this feature this morning. It looks like we DID find some issues, but that the behavior from your video is actually expected.
When Snap to Zero Crossings is enabled, moving the CTI or making a selection should ensure the boundaries are set to the nearest zero-crossing, at least for Mono and Stereo files. (This feature is disabled for files with more than 2 channels.) However, when you zoom in to a level where you can see the individual samples, snapping to zero crossings is overridden since it would prevent accurate editing and may result in wildly unexpected behavior depending on the waveform and zoom level. From what I understand, when you click to position the CTI, there is a 7-pixel span in which the algorithm looks for the nearest zero-crossing. When zoomed out, this can represent a significant number of samples, so the snapping is accurate and works as expected. But when zoomed in to a level where 7 pixels represents only a few pixels (or less) snapping is overridden in favor of accurate editing.
The bugs we found show a tendency to favor the right channel for locating zero crossings in stereo files and in mono files, an off-by-one offset that positions the CTI at the sample interpolation position just before the zero-crossing. This bug is being investigated.