[asukaminato0721]

Summary

Fixes #2071

By let user write these models down in configure.

• I am a human, and I fully understand the code I am submitting.

Test Plan

add test

[meta-codesync]

This pull request has been imported. If you are a Meta employee, you can view this in D123799553. (Because this pull request was imported automatically, there will not be any future comments.)

[asukaminato0721]

todo: this will need manually add configure, is it possible to do some cold start?

[MorganBennetDev]

Not familiar with how the Rust code works, but will this support globs (src/*/models.py) or only explicit paths?

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[rchen152]

@asukaminato0721 thanks for taking another stab at this. Sorry for never getting around to posting design ideas as I promised to on your last PR.

This direction looks promising to me, but AI review turned up a couple potentially blocking correctness issues (1, 2) and a couple more minor issues. I haven't vetted these carefully, so please feel free to push back on anything the AI got wrong:

1. The ThreadState cache loses dependency edges and ignores per-module config. Only the module whose solve fills the cache records the dependency; models whose fields are computed on demand later in that solve hit the cache and record nothing, so their reverse fields can go stale, depending on solve order. In a project with nested configs, one sub-project's list can also leak into another's. answers.get already memoizes, so the fix is to delete the cache.
2. Invalidation never fires. KeyDjangoRelations doesn't override Keyed::try_to_anykey in pyrefly/lib/binding/binding.rs, so diffing a module's results always skips it. Editing a listed module's relations is never reported to the modules that read them, so in the IDE every cross-module reverse field would go stale after an edit, cache or no cache. This needs a fix and an incremental test.
3. (ux weakness) Bad config entries are dropped silently. An entry that can't be resolved, such as a typo, gives no warning.
4. (nit) The merge code is clunky. The four-arm match merging the two Options can be a single chain(...).reduce(ClassSynthesizedFields::combine).

@samwgoldman I'd love to get your thoughts on this change, since you're much more familiar with the state machinery.

[rchen152]

(Forgot to submit a review XD)

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅