Increasing the OP_CODESEPARATOR weight by 520 (p2sh redeemScript size limit) + 40 (stripped txinput size) + 8 (stripped txoutput size) + a few more (overhead for varints) = 572ish bytes should be enough to completely eliminate any vulnerability caused by OP_CODESEPARATOR within P2SH transactions without the need to remove it ever.  I think it is worth attempting to be a bit more clever than such a blunt rule, but it would be much better than eliminating OP_CODESEPARATOR within P2SH entirely.

Remember that the goal isn't to eliminate OP_CODESEPARATOR per se; the goal is to eliminate the vulnerability associated with it.