Hugh, I didn’t know this was a thing. Cool.

This change breaks the resulting tests, namely the assumption tests. Revert at #321246.

Serves to illustrate that a) all changes should be tested properly even if it appears that making them couldn't possibly break anything (I admit, in this case it is a little surprising, but ofborg correctly reports the other tests as changed!) and b) grepping for src = ./. is not enough, as the ./local path would also contain generated.nix. This is maybe a useful hint for work on #301014 that all local paths in nixpkgs will eventually need to be checked.

Sorry. but serious question: What's the correct workflow to catch this. I remember verifying that ofBorg built the tests correctly. Clearly that was not enough, as I didn't know there were more tests affected. My general stance was, that we will see build errors once they reach hydra, (which is a luxury we would have if we merge into haskell-updates, I should insist on that more)...

Yeah, I'd love to learn. I thought I did my due diligence (see the tests listed).

I see after investigation into the breadcrumbs you left in #321246 that the test is:

• tests.haskell.cabalSdist.assumptionLocalHasDirectReference

This test (along with others) were added in #174176 by @roberth.

Specifically, what I'd like to learn is:

1. What is tests.haskell.cabalSdist.assumptionLocalHasDirectReference actually testing? The builtins.unsafeDiscardOutputDependency appears to be load bearing there, and there's no documentation of that function that I can find. What's going on there?!
2. Why adding the default.nix file back in caused the test to pass. How is the presence or absence of the file causing Nix to have different behavior here?

1. What is tests.haskell.cabalSdist.assumptionLocalHasDirectReference actually testing?

A non-essential property of the way Haskell derivations are put together. If the assumption is broken, that suggests that the test may need to be updated, specifically because the more relevant test localHasNoDirectReference may succeed even if buildFromCabalSdist becomes a no-op by mistake.

1. The builtins.unsafeDiscardOutputDependency appears to be load bearing there, and there's no documentation of that function that I can find. What's going on there?!

It turns off the implicit behavior that Nix will build the outputs of a .drv if you refer to it like that. It's a performance optimization, and not really load bearing. It just makes it so that you're not substituting many unnecessary gigabytes.
Can't find the right phrasing right now, so an issue it is

• Missing documentation: unsafeDiscardOutputDependency nix#10965

2. Why adding the default.nix file back in caused the test to pass.

The test just treats ${./local} as a constant. In any test, if you change a single constant literal to a different value, you can expect it to fail. Take for example assert 1 == 1;; change one of the integer literals, and then it fails.
This is so obvious that I guess you might feel offended. Of course it's not obvious in this case, because the literals are in different files, and they don't even look the same, so how could you know.

• ./. in generated.nix
• ./local in the test

Both refer to the same thing. I just didn't communicate my intent.

I've opened #322556, in case it helps.

That explains it. Thanks, Robert.