Checkpoint numpy codec: structured dtypes corrupt silently, object arrays persist raw pointers, and round-tripped arrays are read-only

Finding: In an external report against langgraph-checkpoint 4.2.0 (numpy 2.4.1, ormsgpack 1.12.2, Python 3.13.12, Windows), the JsonPlusSerializer numpy array codec (EXT_NUMPY_ARRAY) reportedly mishandles three kinds of arrays on a dumps_typed/loads_typed round trip. Source: https://github.com/langchain-ai/langgraph/issues/8746. Limitation: this is an external report; KFA did not reproduce it, and no maintainer has triaged it. Reported symptoms in the issue body: (1) structured dtypes come back as an opaque void dtype (|V12) with no exception, so field access no longer works; (2) object arrays write raw in-process pointer bytes into the payload and decode to None without error; (3) restored arrays are read-only, so in-place updates fail with ValueError about a read-only assignment destination. The reporter attributes all three to the codec describing arrays with dtype.str plus a raw buffer. This cause is the reporter's analysis and is not maintainer-confirmed. The reporter also notes the pointer persistence may leak address information to anyone who can read checkpoint storage, but judges it not a reportable vulnerability. That assessment is unvalidated. A follow-up comment (https://github.com/langchain-ai/langgraph/issues/8746#issuecomment-6030145178) reports that the proposed repair in closed, unmerged PR #8747 mishandles aligned structured dtypes. A helper-level check on four extracted functions reportedly turned names ('a','b') into ('a','f1','b'), lost the alignment flag, and made a structured-to-unstructured conversion raise DTypePromotionError. That check did not run the full serializer, a checkpointer or the repo tests. The issue author then agreed that gapped layouts break the same way and said they changed their fork branch to store structured dtypes in numpy's dict form (https://github.com/langchain-ai/langgraph/issues/8746#issuecomment-6041558362). They report 179 passing tests in libs/checkpoint, excluding Redis tests. Only the author says so; it was not independently verified. Fix status: proposed only. The sources show no merged commit and no released fix, and #8747 is described as closed pending issue assignment. No maintainer response appears in the supplied comments. A user who stores structured, object-dtype or later-mutated numpy arrays in checkpoints could check whether restored state matches the original, including dtype layout and writability. The sources do not provide a verified workaround. The comment record is a public initial head/tail view and may omit middle comments.