Testnet is broken because too much power was puten there (thanks Xelister 1 and Artfortz 8 of 5970's

It will take apparently months (with 1 5970) to reset diff to be usable by CPUs.

Is this design flaw? I think so.

Solution: lower diff also each say 72 hours passing with too few (or none!) blocks.

Also disallow overtaking network and forcing switch to own version of chain - with like totally other history for last 1 week, or 1 month, or even since the begining!
Solution: use more anchors (e.g. fixed points in chain on which everyone agrees).
Not only in source code but also distribute them. Say each 2 days?

Adjusting the testnet difficulty every, oh, say 100 blocks (instead of 2016) seems like a good idea to me.

That's a trivial patch; I nominate ArtForz and Xelister to write it and submit a pull request.

Locking in the testnet block chain-- not so much.  I don't think it is worth the effort to either come up with a mechanism for distributing the lock-in points or manually put lock-in points in the source code.

Why do I feel that at some point we'll drop the main network and stick with the test network

Testnet can be reset within 2 days by overtaking the blockchain. This will reset all balances. I might actually do that this week. Final difficulty will be ~ 0.5 after reset.

It's a bug in test net because no-one (comparatively) was using it.

It's impossible in live, because by the time governments care and devote their super-secret NSA SHA-256 bruteforce clusters, they still won't make that much of a dent

4x rule was considered and actually makes overtaking easier. You can create 5 (iirc) blocks per second (implicit limit in AcceptBlock()). Without the 4x rule difficulty would rise to too high after the first 2016 blocks. With the rule, it will rise to 4 and we can still generate the next 2016 blocks in a short time.