[asukaminato0721]

Summary

Fixes #5322

Django fields now accept F(), Value(), and combined expressions while preserving their read types.

• 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 D127050295. (Because this pull request was imported automatically, there will not be any future comments.)

[stroxler]

Thanks! I think this causes a couple of behavior regressions based on a round of review:

Ordinary assignments lose nullable-field narrowing

[class_field.rs:5550–5563](https://github.com/meta/pyrefly/blob/d58a18a517f7168cbb0b551c48ba837b9c7735ed/pyrefly/lib/alt/class/class_field.rs#L5550-L5563)

The new setter returns narrow: false for every Django-field assignment—not just expressions. For an IntegerField(null=True):

book.pages = 1
return book.pages + 1

This now reports an unsupported operation involving None. Nullable strings and foreign keys regress similarly. It also breaks the common “if missing, assign a default, then use it” pattern, consistent with the added nullable-access errors in the primer report.

Fix: preserve assignment narrowing when the RHS fits the read type; suppress it for expressions or coercible values outside that type.

Choice-restricted fields accept values contradicting their read type

[django.rs:325–332](https://github.com/meta/pyrefly/blob/d58a18a517f7168cbb0b551c48ba837b9c7735ed/pyrefly/lib/alt/class/django.rs#L325-L332)

Only the getter is narrowed by narrow_charfield_choices. The setter remains str | int | Combinable, so this becomes accepted:

# suit has choices "A" and "B"
card.suit = "INVALID"
assert_type(card.suit, Literal["A", "B"])

Previously the assignment was rejected. This makes Pyrefly’s existing finite-choice guarantee inconsistent; it is not a claim that Django validates choices during assignment.

Fix: retain choice restrictions for ordinary writes while permitting expressions, or deliberately revise the finite-choice read guarantee.

Expression assignments still fail inside model methods

[class_field.rs:5550–5558](https://github.com/meta/pyrefly/blob/d58a18a517f7168cbb0b551c48ba837b9c7735ed/pyrefly/lib/alt/class/class_field.rs#L5550-L5558)

The new branch requires instance_class, which attribute resolution does not supply for implicit self or bounded type variables:

class Counter(models.Model):
    count = models.IntegerField()

    def bump(self) -> None:
        self.count = F("count") + 1  # Still rejected

An explicitly annotated self: Counter works; def update[T: Counter](counter: T) still fails. This is an incomplete fix rather than a new regression.

Fix: carry the write type through attribute lookup instead of recovering it through a second Django-specific member lookup. That also avoids duplicate lookup and instantiation.

[stroxler]

Review automatically exported from Phabricator review in Meta.

[stroxler]

It's looking pretty good! I was going to land with fix-forwards, but I think it already needs a rebase; the codebase is churning pretty fast now as we try to clear the backlog.

Could you rebase, and also address these issues that I was handling in my fix-forward:

• Respect explicit annotations. An inherited Django field’s permissive setter must not make self.count: int = "123" valid. Annotated assignments should retain normal type and qualifier checks.
• Preserve override compatibility. Django coercion types shouldn’t create new override errors for EmailField/TextField replacing an abstract CharField, or for overriding annotations/properties. Keep read-type invariance and explicit parent-property setter requirements.
• Handle intersection receivers. Fold distinct write types instead of falling back to one receiver member. Writes must satisfy every member, and read-only/Final restrictions must survive.
• Keep ordinary attributes cheap. Don’t allocate separate Django read/write collections for unions where nothing transforms.
• Improve diagnostics. When checking a field against a property setter, report the effective write type, not the read type.

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

============================================================
SUMMARY
============================================================
Total: +0 new errors, -1 fixed errors
By preset: +0/-1 (default), +0/-1 (strict)

Projects with changes (1):
  zulip: +0 -1
============================================================

FULL DIFF DETAILS
------------------------------------------------------------

zulip (https://github.com/zulip/zulip)
- ERROR zerver/lib/test_classes.py:2859:28-44: `Literal['6cde5f7a-1f7e-4978-9716-49f69ebfc9fe']` is not assignable to attribute `uuid` with type `UUID` [bad-assignment]

[meta-codesync]

This pull request has been merged in 9b870b9.