Fix no-op for initializeObject() and isUninitializedObject() when object is not a mapped entity

[SherinBloemendaal]

Q	A
Type	bug
Fixed issues	#12172

Summary

Supersedes #12173, which was auto-closed as stale and can no longer be reopened because its base branch 3.6.x has been deleted. Same change, rebased onto 3.7.x.

Fix initializeObject() and isUninitializedObject() to be a no-op for non-managed objects when native lazy objects are enabled.

With isNativeLazyObjectsEnabled(), both methods call getClassMetadata($obj::class) unconditionally, so passing a non-entity (a DTO, stdClass, …) throws a MappingException instead of the no-op the ObjectManager interface documents. This breaks e.g. Symfony's UniqueEntityValidator on DTOs with entityClass, which calls $em->initializeObject() on association field values.

This fix wraps the getClassMetadata() calls in try/catch blocks to handle non-entity objects gracefully while preserving functionality for valid Doctrine entities (including detached ones after clear()).

I went with the try/catch approach since isInIdentityMap() also internally calls getClassMetadata(), but I'm open to better solutions if there are any.

Changes:

• UnitOfWork::initializeObject(): Add try/catch around metadata access
• UnitOfWork::isUninitializedObject(): Add try/catch around metadata access
• GH12172Test: functional test covering both methods for non-entities and for real lazy entities

[greg0ire]

A try with an empty catch is an anti pattern I think. Have you considered using AbstractClassMetadataFactory::isTransient() instead?

[SherinBloemendaal]

Good point, thanks. Switched both methods to an isTransient() guard and dropped the try/catch (5c03f84). The MappingDriverChain test cases stay in GH12172Test, since that driver returns true from isTransient() for a class outside every registered namespace — which is the Symfony case that triggered this.

One note: AbstractClassMetadataFactory::isTransient() always goes to the driver, so for an already-loaded entity class this is a ReflectionClass + attribute read per call, on the computeAssociationChanges() path. Probably fine, but if you'd rather avoid it I can add a hasMetadataFor() fast path — or that could be a small follow-up in doctrine/persistence (check loadedMetadata before asking the driver).

[greg0ire]

I think a follow-up in doctrine/persistence would make sense. Actually, you could send it right now, it's orthogonal to this PR, isn't it?

[SherinBloemendaal]

I tried the bare isTransient() first and it broke ValueObjectsTest::testPartialDqlOnEmbeddedObjectsField: an embeddable loaded through a partial query is a lazy ghost, but isTransient() reports it as transient, since by contract it is only false for entities and mapped superclasses. So isUninitializedObject() returned false for an uninitialized embeddable ghost.

865fbf2 keeps isTransient() but checks hasMetadataFor() first. If this manager has already loaded metadata for the class, that covers embeddables (the ghost could not exist otherwise) and makes the common case an array lookup, which also takes care of the reflection cost. isTransient() remains as the fallback for an entity class this manager has not loaded yet.

I also drafted the doctrine/persistence change I floated earlier and then dropped it. Short-circuiting isTransient() on loadedMetadata would make it return false for embeddables (and ODM embedded documents) once loaded, which contradicts its documented contract and makes the answer depend on load order. The abstract factory cannot tell a mapped superclass from an embeddable via isEntity(), so I do not see a safe generic version. Sorry for the detour.

[greg0ire]

Thanks @SherinBloemendaal !