v3 already has significant enhancements over a "pure ARX" design or even RC4 :

• The LUT is NOT a byte permutation. It's a bit permutation. This makes the output useless to recover the input/index.
• The index addition for A and B is chained/merged hence co-dependent. Getting one index from the output does not directly provide the other index, particularly since A is XORed by the LFSR.
• The LUT is an expansion table and only one half is output, the other is fed back.

But it's not perfect yet.

https://en.wikipedia.org/wiki/Rotational_cryptanalysis describes the class of cyphers based on the Add/Rotate/XOR operators, ARX in short. It is the basis of Salsa20, Speck, ChaCha20, ThreeFish, Blake, Skein, Cubehash...

Only the (truncated carry) Add is "not linear", so the cypher requires a lot of rounds. It also hints how its cryptanalysis works, in particular because XOR, ADD and ROT are reversible. Some other cyphers combine non-reversible operators : OR and AND. v4 would benefit from this.

Since the design of #PEAC I have a fondness for End-Around-Carry, which wipes the LSB computation that degenerates to a simple XOR, that is easy to correlate. So adding a carry-in and carry-out to the design will strengthen it. That's a second enhancement for v4.

A third easy enhancement is the replacement of the lousy LFSR8.