Recommended patterns when using resolvers

From our discussion related to #4967, my understanding is that while Relay resolvers can be used in many different ways, some of them are more idiomatic and have a lower risk of gaps or weird corner cases. I'd like to ask a few questions related to that, and I'd be happy to contribute to the docs afterwards so that this knowledge is captured somewhere 🙂

1. Should all client-defined types be implemented either as "strong" or "weak" types, and "output" types (i.e., types declared in client schema extensions) should be avoided?
2. When creating a resolver for a field that corresponds to a one-to-many edge to a strong type (i.e., `SomeType.myField: [MyStrongType]`), should the resolver return a list of `{ id: DataID }` objects? And if the resolver is live, does it require the strong type resolver to also be live?
3. When creating a live resolver that does some expensive work (e.g., doing a REST API call), should this expensive work be done in the body of the resolver or in the `subscribe` function?

Start by reviewing the discussion in #4967 and the three resolver questions in this issue. The issue names no documentation file or test; done means resolving the questions and capturing the agreed recommended patterns in the project documentation.

Written by the indexing model from the issue text.