Data-aggregator-mcp

Added

- A search hit ranked on another copy of the same deposit now names that copy in
`ranked_by_copy`. The hit shown is the most fetchable copy, and the rank comes from the
best-described one, so a DataONE copy without a description could top a search on its
DataCite twin's species list with nothing to say why.

Added

- `massive`, MassIVE searched natively. The round-3 coral study MSV000101299 names its
marine heatwave only in its description; OmicsDI's copy of MassIVE lags, and DataCite
holds every MassIVE DOI with the title alone, so no source found it for a heat-stress
search. MassIVE's own search reads titles, descriptions, keywords and species, but ORs
words and matches inside them ("heat" finds "wheat"), so the rarest word's datasets are
read and a dataset is kept when it names every word. A dataset MassIVE, DataCite and
OmicsDI all return is listed once. `resolve` adds the DOI and FTP location.
Discovery-only.

Fixed

- A JSON answer that is not UTF-8 failed as an outage. MassIVE labels its answers
ISO-8859-1 and sends that, so any answer holding a "±" was unreadable. A body now falls
back to the charset its server declares.
- On Python 3.14, every `operate` query stopped at the wall-clock limit logged
"InterruptException exception in shielded future": from 3.14 a cancelled `asyncio.shield`
logs the stopped query's error itself. The query is now awaited without `shield`, and CI
runs 3.14.

Added

- `mwb`, the Metabolomics Workbench searched natively. The Workbench was reached only
through OmicsDI, whose copy stops at ST002490 while the Workbench is at ST005288, so 2,455
newer studies could not be found; the round-3 benchmark's coral heat-stress studies
ST004143 and ST004329 were two of them. The Workbench searches titles only, so each word
of a query is its own title search, and a study is kept when its title, species,
analysis type and institute name every word. A study both sources return ("diabetes":
23) is listed once, and OmicsDI records now carry their accession. Discovery-only.

Fixed

- CELLxGENE read a query as raw words split on spaces: "tendon single-nucleus human" found
nothing although a collection is named "Single nucleus and spatial transcriptomic
profiling of human healthy hamstring tendon", "A OR B" needed the word "or", and a quoted
phrase kept its quotes. It now reads a query as the rest of the server does: hyphens and
punctuation as spaces, a quoted phrase whole, no operators or stop words. Over the 398
live collections and 40 queries, no match was lost and 191 were added ("T-cell",
"single-cell", "alzheimer's"). When a search of two or more words empties CELLxGENE,
`errors["all_words"]` now names it, as it does GEO and OmicsDI.

Fixed

- `organism="water bear"` resolved to a bear (*Ursus* sp.) and filtered the search to it,
and `organism="tardigrade"` resolved to nothing: NCBI Taxonomy dropped the words it
could not find, and names Tardigrada only in the plural. An organism is now looked up
as a whole name, singular or plural; both resolve to Tardigrada. Of 28 common names
checked, no other resolution changed.

Changed

- GEO, SRA and BioProject are sent each plain word of a query with its plural, as
DataCite, Zenodo and four other sources already were. NCBI maps some plurals, never the
singular, to a taxon, so "tardigrades" also finds Tardigrada records that never name the
word: SRA "tardigrade anhydrobiosis" found 4 runs and none of the *Ramazzottius
varieornatus* study PRJDB2359; the 33 runs the plural adds hold all six. They are asked
for as a separate stream beside the query as written, because NCBI lists records newest
first, and in one query they pushed the runs naming both words past the first page. A
cursor from an earlier version continues without the extra stream.

Changed

- Search ranks a data deposit (dataset, sequencing run, study) before a paper, figure
or tool among hits that match the search equally. DataCite and Zenodo title matches
are now asked for in two parts, deposits first. For "snow leopard", figshare datasets
that papers with the same words in their titles had pushed to 213th or past 240th now
rank 66th to 99th, and the first 240 hits hold 31 figshare records instead of 9.
- A search shows each deposit once: copies and versions a record links to (figshare's
`.v1` DOI and its unversioned twin, the versions of one Zenodo concept) and the runs of
one sequencing study collapse to one hit, the most fetchable copy kept. Zenodo records
now link their concept DOI (`is_version_of`), as DataCite's copies of them already did.
- BioStudies `S-EPMC` records, Europe PMC's imports of papers with their supplementary
files, are `kind="publication"`, not `"study"`. Page 1 of "snow leopard" (50 hits) had
spent 7 slots on runs of one study, 6 on extra copies and 16 on these papers; it now
holds 36 answer-key studies instead of 25, including the four it was missing.