Reviewed PR 356 from merge base d2ab806c through HEAD, focusing on the vxData SDK storage namespace, transfer engine, API transport, and Direct S3 adapter.
Reported prioritized architecture suggestions to the parent reviewer without editing the monorepo.
Recommended keeping wire policy inside the transfer engine, moving multipart presign batching into the transport layer, typing the engine against narrow control-plane interfaces, and reusing batched uploads for DataFrame convenience paths.
Marked generic transfer-task hierarchies, upload/download state inheritance, and a public multipart planning API as over-abstraction.
Lessons Learned: Pitfalls
The general vxData schema guide gave useful SDK context but did not cover the new multipart branch. Direct inspection of the merge-base diff and focused upload tests was required.
A private test seam (StorageNamespace._wire) currently shapes production ownership, which makes the documented claim that request shapes are not caller-tunable weaker than it appears.
Lessons Learned: Improvements
Future SDK architecture reviews should compare where analogous download and upload policies live. Here that comparison quickly exposed multipart batching in the wrong layer.
Review convenience methods as well as the main primitive. The main upload path batches work, while DataFrame and referenced-file helpers still reduce it to sequential one-file calls.
