LOCAL COMMIT ONLY -- this remote is consortium-owned and is never pushed from here. These are handover documents for the owilix maintainer. open_asks_against_5.10.8: most of the 2026-08-07 improvement brief has shipped. Verified live against the deployed server rather than from release notes -- Part 1 (v5.9.0), Part 2 (`_index.parquet`, `_collection_summary.json`, `remote reindex`), Part 4a (`remote verify` + the three verification endpoints) and Part 4b (`partial`/`failed` on `/remote/ls`) are all closed. Written so nobody re-implements them and so the genuinely open items are not buried. datacite_dropped_by_remote_ls: `_dataset_record()` projects `ds.metadata` from 38 keys to 11, discarding creators, abstract, publisher, licence and related identifiers one function before the response. No response model is involved -- `/remote/ls` is an untyped object -- so this is one helper. Two traps documented, because the obvious fix hits both: - there is no nested `datacite` key by the time a record reaches `Dataset.metadata`; the fields are flattened at the top level; - `metadata.get(...)` returns a *display rendering* -- `creators` is 'Lukas Gienapp+7', the abstract is ellipsis-truncated -- so adding `md.get("creators")` to the record would look correct and ship lossy data. Plus a small mapping-contract defect found on the way: `DatasetMetadata.keys()` advertises keys `__getitem__` raises KeyError on, so `dict(md)` fails on exactly the records in question. Presumably why `_dataset_record` hand-picks with `.get()`. Also measured and carried forward: `/health` is still a literal with no readiness split, and pagination still re-walks the backend per page (offset=1800 costs the same 7.3s as offset=0, and an immediate repeat is not cached). Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_019nPNpEG2QRnmjxxn5LeEb7