Imported placeholder names are lost before dynamic class hooks run #22326

An imported class that is still a PlaceholderNode loses its qualified name on the NameExpr. If it is the receiver of a dynamic-class factory, mypy can miss the factory's plugin hook during an import cycle. The resulting generated base is rejected as a variable, and methods inherited through it become Any.

This is reproducible on master at 7da2058491422897ecdbf61261d2540bb10099a0 with Python 3.14.6, without Django or any third-party plugin.

The standalone regression uses the existing dyn_class_from_method test plugin and this import graph:

models -> managers -> base_managers -> models

base_managers:
    BaseManager derives from QuerySet.from_queryset(BaseQuerySet)
managers:
    ChildQuerySet derives from BaseQuerySet
    ChildManager derives from BaseManager.from_queryset(ChildQuerySet)
models:
    Row.objects = ChildManager()

The complete regression and small test plugin are available in the proposed fix. Copying only those test changes onto the upstream revision reproduces the failure with:

python -m pytest -n0 mypy/test/testcheck.py -k testDynamicClassHookFromImportedPlaceholder

Expected results:

main:4: note: Revealed type is "builtins.int"
main:5: note: Revealed type is "builtins.str"
main:6: error: "ChildManager" has no attribute "missing"

Actual results on upstream master:

tmp/managers.py:9: error: Variable "managers._ChildManager" is not valid as a type
tmp/managers.py:9: error: Invalid base class "_ChildManager"
main:4: note: Revealed type is "Any"
main:5: note: Revealed type is "Any"

Assigning expr.fullname = sym.fullname or "" in the placeholder branch of SemanticAnalyzer.bind_name_expr is sufficient to make the regression pass. The full custom-plugin test file passes with that change (65 cases).

This report and patch were prepared with AI assistance from a failure encountered in downstream Django manager typing, reduced to the ordinary mypy test plugin above.