ConvertTo-Json serializes in-depth ... properties that are actually rich objects become nested dictionaries instead of simple strings. For instance, the "Parent" which is a Folder object. Recursively, up to the depth allowed.
Anyway, when you convert back, you don't get actual FileInfo and DirectoryInfo objects, you get dynamic property bags. So those complex properties remain dictionaries ... which ConvertTo-CSV turns into strings in a very different way.
If you specify -Depth 1 when you ConvertTo-Json ... the resulting output would be very similar.
To clean up the image after you're done:
docker rmi trevorsullivan/powershell
For your first question: you have converted types. When you did a ConvertTo-Json, you serialized the System.IO.DirectoryInfo type into JSON compatible data. When you ConvertFrom-Json you have a PCSCustomObject.
The conversion to CSV is different for the System.IO.DirectoryInfo than the PSCustomObject. This has to do with the sub-types (converting from objects into serializable JSON objects). It also appears to handle time strings differently as well.
(ls | % Name) -match '_'
The fact that the file's name is its default string conversion can be pretty annoying when dealing with results, e.g. from gci -recurse where it's quiet important to retain the path to the file. So, careful with such things. On Stack Overflow I tend to scold people who reduce PowerShell objects to strings without good reason ;-)
Edit oooo, and all the linux stuff is available within it too. So you can run this to just augment your linux experience.