Batch-split-on-OOM retry loop: re-queuing the remainder in the same order re-forms the failing batch

A common bug in batched loaders that split a batch when it exceeds a memory cap: on overflow, the code peels off the first item as a single, puts the other N-1 back at the head of the queue in the same order, and retries. The next iteration takes N items from the head again, which re-forms nearly the same oversized batch, and it fails again. Every heavy item then pays one failed full-batch load, and the memory peaks come from those failed loads. In one run this was about 72% of all worker time, at roughly 12x the normal per-item cost. Fixes: on overflow, process the rest of the failed batch as singles, or bisect it (halve and recurse), and never let the next batch re-form from the same items. Log split events with batch size and item IDs so the pattern shows up in the logs, and test with a fixture that puts a heavy item mid-batch.