Core problem: If contracts can be upgraded while in use, everything reduces to a game of Nomic.
A contract should present an interface that includes a declaration of its behavior. The declared behavior should be well defined, and if a bug in the implementation is discovered, the contract should be updatable to fix the bug. There could even be futures expressing the probability that a contract will be found to have a bug.
The willingness of participants to use a contract would depend on the chances that something about the interface is bug-prone, untested, etc.
As long as bugs do not result in reversal of money flow, the incentives seem to align properly toward a well-defined approach for declarative contracts.
The first part, a "declaration of...[future] behaviour," is basically a normal contract.
There ought to be a way to have highly vetted primitives. In meatspace legalese, boilerplate words and phrases are the closest we get to this... once a contract (or open source license, etc.) has been through litigation, its vulnerabilities become better known.
If a dispute gets decided the "wrong" way because a few clarifying words were absent, the contract is modified and future deals use the new contract.