Same protein P is coded by two different sequences, A and B. The "redundancy" observation is just saying that when you find some P molecules in the organism, you can't tell which were coded by A and which by B - that information is, literally, lost in translation. Yes, it's not really lost - it's smeared over the "global state", as DNA sequences are physical objects and A and B have slightly different resource costs, but that's still far from being able to modify DNA based on information from proteins.
FWIW, "redundancy" does have some different interesting functions in nature. Sequences like "AAAA" "ABAB" and "BBBB" are (in this example) equivalent under normal read, but if you shift the reading window one nucleotide to the left or right, those two sequences could code for something else entirely. AFAIK some bacteria exploit that. But again, this is still not a protein->DNA information transfer.
The argument goes like this;
Suppose you built a library of chemical reactions, to generate a DNA codon from an amino acid, and did this for all of the different amino acids. Since two different DNA sequences comple to the same protein, you could take sequence A, perform transcription and generate a protein P, and then construct a different DNA sequence B.
Thus we have "genetic information" extracted from a protein and stored in DNA, violating the Dogma... in an extremely cheesy way, since it's obviously the same "information" you had at the start, just encoded differently, and also it's not different in any meaningful sense from doing the same reverse engineering on non-logically-redundant DNA.