That way you could get feedback from other devs, and pushing commits to your PR would trigger CI runs so you could be made aware of any build failures you were causing. (The builds took hours, sometimes, so running the full test suite locally was impractical)
We used Github but I'm sure it would work in a lot of workflows.
So what my team does when we want to prototype something is we create a change and then push separate patch sets for different solutions (there is a nice diff tool to view differences between patch sets). The CI runs for each patch set and we then discuss the merits of each approach and "abandon" (reject) the change and use it for reference going forward.
So you can directly use them to show a what-if PR. I did; it worked, that is, sparked a discussion and led us to designing things in a better way.
My objection to using PRs for this is that people think the point of a PR is to merge the code, and people tend to nitpick the code during code-review with the goal of making it clean enough to merge.
The idea of a separate RFC object is that, unlike a PR, you literally cannot merge an RFC, so there's no temptation to nitpick, or really to talk about anything other than the design. It much more closely mimics the social mores of a mailing-list thread discussing a code snippet.
Also, being able to explicitly track Approved and Rejected RFCs on a system level would be nice, to know what discussion needs to be referenced when doing the final implementation. If you just used "what-if" PRs, both the chosen and not-chosen designs' PRs would just end up in the Closed state, and would show up equally in search. Ideally, Rejected RFCs would be filtered out of search by default.
I kind of wish that there was a type of thing like a PR, but marked in a way where you couldn't actually merge it. It'd still get built by CI, if such a thing was configured; but the point of the submission of such a "proposal with prototype" object would be to discuss whether the design represented by the prototype implementation is a design worth going with.
The actions applicable to such objects would be "accept and close" or "reject and close." You'd be able to have several of these objects that live under a given issue (i.e. several potential solution-designs to the same problem), and—as long as the issue has at least one proposal-with-prototype object under it—the issue would be in a "pending" state until one such proposal object was Approved, and Approving one proposal object would Reject the others.
The point of this would be to replicate the thing that people go through with design discussions on mailing lists when they send in code samples to explain their designs—but with those code samples being working, buildable code, such that the properties of the design proposal can be tested against the current implementation and against any alternative proposals.
You could call these objects "RFCs" :)