Usually the person parsing the CSV data doesn't have control over the way the data gets written. If he did, he would probably prefer something like protocol buffers. CSV is the lowest common denominator, so it's a useful format for exchanging data between different organizations that are producing and consuming the data.

https://github.com/dbro/csvquote is a small and fast script that can replace ambiguous separators (commas and newlines, for example) inside quoted fields, so that other text tools can work with a simple grammar. After that work is done, the ambiguous commas inside quoted fields get restored. I wrote it to use unix shell tools like cut, awk, ... with CSV files containing millions of records.

I used that strategy for parsing gigabytes of CSVs containing arbitrary natural language from the web - try to get these files fixed, or figure out a grammar for gigabytes of fuzzy data...

My approach never failed for me, so telling me that my strategy does not work is a strong claim, where it reliably did the job for me.

Your examples are all valid, but what you are describing are theoretical attacks on the method, while the method works in almost all cases in practice. We are talking about two different viewpoints: dealing with large amounts of messy data on one hand and parser theory in an ideal cosmos on the other hand.