Firstly, you may want to apply COMPRESSION_DELAY option for those COLUMNSTORE INDEXES to reduce the fragmentation as deleting compressed rows would cause significant fragmentation and the storage remain used even for deleted rows, eventually there is negative impact on the performance.
MS Docs - Performance tip
While the expected ...
In a word, NO!
You should not breach Codd's Rules unless you have no choice but to do so.
I fail to see what performance gains you will obtain by having to parse strings over having an Associative Entity (AKA joining, bridging tables or mapping - amongst others) and using SQL JOINs.
Most importantly from an RDBMS perspective, this allows you to ...
Correct, you don't have to.
For inserts be sure that ETL is saving data directly into compressed rowgroups, not the delta store.
If you want to update or delete data you need to be aware of the internal implementation of the updates on the column store. During DML operations data in a compressed column store rowgroup are never deleted or modified but ...