I just had a great discussion with a developer who urged me to write a "big picture" technical post about BIP 16. So:
First, I think a good design approach is to be clear about what you are trying to accomplish and think about what the ideal solution, if there were no constraints like backwards compatibility, would look like.
The big picture goal has always been: short, multisignature bitcoin addresses (BIP 13).
The ideal solution would be to split scriptSig/scriptPubKey into three parts:
signatures, redemption script, and redemption script hash.
The sender would just provide the script hash, the receiver would provide the script and signatures to sign them over to somebody else.
Ideally, the redemption script hash would be include a version or 'hash type' byte, so in the future if RIPEMD160(SHA256()) was ever considered insecure a smooth upgrade could happen.
That's the ideal solution. I think all bitcoin transactions should have been done that way from the start, but it is what it is. Now we have to compromise, because one of the design constraints is backwards compatibility-- we are not going to replace scriptSig/scriptPubKey with something else and require everybody to upgrade all at once.
OP_EVAL tried to do too much, in my opinion. It enabled all sorts of nifty things, but we made the mistake of losing sight of what we were trying to accomplish.
BIP 16, in my view, meets the goal and (importantly!) does nothing more. I think of it as implementing the ideal three-way split in as simple and safe way as possible:
signatures: all the items except the last pushed onto the stack by the scriptSig
redemption script: the last item pushed onto the stack by the scriptSig
redemption script hash: the scriptPubKey
It is pretty darn close to what I think would be the ideal solution. It even has a byte at the beginning of the redemption script hash that specifies what hash type to use (OP_HASH160) !
That's all why I like BIP16 better than OP_EVAL. I've written quite a lot here on the details of why I prefer BIP 16 to BIP 17, but, fundamentally, I believe that BIP 16 is a more conservative solution that is less likely to have an "darn, I didn't see that coming" bug.