Less test generation

[jneem]

This switches our path-based tests from test_generator's compile-time generation to runtime generation using libtest_mimic.

This has a couple of benefits:

• new test files are automatically picked up without having to trigger a rebuild some other way (for example, in cargo test; echo '{' > cli/tests/snapshot/inputs/pretty/asdf.ncl; cargo test the second test succeeds because the new snapshot doesn't get picked up
• modifying existing test files doesn't trigger any recompilation

It also has a couple of downsides:

• it's different from the usual test registration story, and the custom runner doesn't automatically recognize #[test]-annotated tests.
• it doesn't capture stdout by default. That's apparently unfixable because of some black magic in the rust standard library. Workarounds include cargo nextest, or just living with messy test output.

While making these changes, I discovered some core integration tests that weren't actually being run, so I moved them to a place where they'd be picked up automatically.

[yannham]

it's different from the usual test registration story, and the custom runner doesn't automatically recognize #[test]-annotated tests.

What does this mean in practice?

[jneem]

So for example in lsp/nls/tests/main.rs we were running a bunch of file-based snapshot tests but also a couple of special-cased tests. After switching to libtest-mimic, those special-cased tests were no longer registered by the #[test] macro, so I had to either register them manually or move them to another test file that uses the default harness (which is what I did).

[yannham]

I see. This is a potential footgun, so I suppose it must be properly documented that one should not mix custom-harness-based tests and normal tests. The stdout part isn't a showstopper, but it is a degradation of the quality of life.

[jneem]

Yeah, overall I'm not totally sold on this approach. It looks like there's some interest upstream in improving the testing situation, so maybe we can just wait for that.

[jneem]

I'll close this for now and revisit if the built-in test situation improves.