jsonplus cannot deserialize generic pydantic(v2) types correctly

Finding: In an external report (not reproduced by KFA), JsonPlusSerializer in langgraph-checkpoint reportedly fails to round-trip a parametrized pydantic v2 generic such as MyGeneric[MyModel]: the reporter says a dict is loaded instead of the model instance. The report lists pydantic 2.11.7, langchain_core 0.3.68, Python 3.13.5 and Darwin (macOS) [issue](https://github.com/langchain-ai/langgraph/issues/6102). A later commenter, not the original reporter, says they reproduced the problem on main commit 7daa3ab; this is a self-report from an AI-assisted commenter and is unverified by KFA [comment](https://github.com/langchain-ai/langgraph/issues/6102#issuecomment-5847016819). Symptom: the reporter's test calls dumps_typed (msgpack) and then loads_typed, then asserts the original instance equals the result. The assertion fails. The reporter's description says a dict object was loaded, and the failure output compares the original instance against {'inner': {'hello': 'hello'}} [issue](https://github.com/langchain-ai/langgraph/issues/6102). The issue shows only the assertion failure output; whether the round trip itself raised or logged anything is not shown there. Condition: the reporter says it works only when the specialized type is also defined as a module-level name (MyNewModel = MyGeneric[MyModel]) [issue](https://github.com/langchain-ai/langgraph/issues/6102). Contributor analysis (hypothesis, not maintainer-confirmed): the commenter says the serializer writes the class name 'Box[Inner]' and the decoder looks that name up on the module, so the lookup fails when no such attribute exists. The same commenter, describing behaviour on main, says there is a warning but no error (the warning text is not shown), and that in strict mode (LANGGRAPH_STRICT_MSGPACK=true) allowlisting the origin, the argument, or the mangled name still yields a dict [comment](https://github.com/langchain-ai/langgraph/issues/6102#issuecomment-5847016819). Fix status: not fixed in any source supplied. Commenters describe PRs #9096 and #9203 that serialize the generic origin plus type args and check the allowlist on all of them. They say legacy payloads still load and that both PRs were auto-closed by an assignment-gate workflow [comment](https://github.com/langchain-ai/langgraph/issues/6102#issuecomment-5925420282) [comment](https://github.com/langchain-ai/langgraph/issues/6102#issuecomment-5990543965). No merge, fix commit or release is evidenced. The issue is open, and no maintainer has responded in the supplied comments. Limitations: all evidence is external reports from commenters with no maintainer association. KFA did not reproduce the problem or run the PR tests. The PRs themselves were not supplied. The affected released version is unknown. The comment timeline may omit middle comments. Actionable diagnosis if you hit this: check whether your state contains a parametrized pydantic generic created inside a function or in another module, and whether you got a dict back from a checkpoint. The module-level alias is the only workaround described in the supplied sources, and KFA has not verified it. One commenter says they worked around the problem without describing how [comment](https://github.com/langchain-ai/langgraph/issues/6102#issuecomment-5857539557), so that workaround is unknown.