Late/optional decoding?

I have a few interrelated questions. Some might even be blasphemous for lack of in-depth knowledge of the wire protocol on my part.

1. Is there a way to delay decoding results until actually looking at data and interpreting it would become necessary?
2. Similarly, would it be possible to take undecoded "records" (not in the strict field-to-value map sense of asyncpg records) and pass them as-is to an upsert without having to re-encode them, so long as the bound arguments correspond to those returned by the previous query?
3. Would that potentially avoid the need to issue type introspection queries?
4. If not as of today, would this become easier once @vmarkovtsev's asyncpg-rkt fork gets merged back upstream?
5. Otherwise, at the very least, would it be possible to decode timestamps to something closer to their internal postgres representation - for example mapping them to the Python's decimal type? I personally find that python's datetime is terrible and Postgres timestamps can fall outside of the range that can be represented via datetime. Should such a decimal codec come with asyncpg, or should everyone write their own if and only if they think they need it?

Thank you & happy new year!

No files, tests, or entry points are named. Start by tracing asyncpg's decoding, type-introspection, record forwarding, and timestamp codec paths, then determine whether these requests can be scoped independently; done would require a maintainer-approved design or a clearly bounded implementation plan.