UFAL/Add CCMM 1.1.0 OAI-PMH crosswalk (ccmm-xml metadataPrefix) - #1262

…ields it dropped

An adversarial audit over all 2628 live LINDAT records found 30 defects. They fall into three
groups, and the first is the one that matters: the feed was stating things the source does not
support.

False statements, now fixed:
- access_rights was taken from others/access-status alone. DefaultAccessStatusHelper derives that
  from the primary bitstream's READ policies, which on a CLARIN repository does not say whether
  the resource can be downloaded: the gate is the licence category in dc.rights.label plus
  others/restrictedAccess. 1216 of 2628 records carried a wrong statement - 1161 anonymously
  downloadable datasets published as "restricted access", 55 login-gated ones as "open access".
  Verified against the live repository: 12/12 sampled "restricted" records return HTTP 206 to an
  anonymous ranged GET, 11/12 "open" ACA/RES records return 401.
- Every date_type label was the register's concept id ("Issued") rather than its prefLabel
  ("Date Issued") - 10512 labels, none correct.
- conforms_to_standard identified the profile by this stylesheet's own XML namespace under the
  label "CCMM 1.1", a string that appears nowhere in the CCMM release.
- license/label carried the CLARIN access category (PUB/ACA/RES) instead of the licence name,
  which was demoted to a description. 1932 records.
- Hierarchical vocabulary paths ("People::Masaryk ...") were published verbatim as subject titles;
  the leaf is now the title and the path a classification_code.
- Agent type was hard-coded per source field, so film studios and ANVL placeholders were published
  as people. Names are now typed from the value, and placeholders produce no agent at all.
- publication_year took min() over the approximate years even when the record stated an exact
  issue date; a comma enumeration ("1920, 1932") became a continuous interval; a "0000" issue date
  fell through to the ingest timestamp; Created ignored the creation year the record states.
- RelationType/Replaces and IsReplacedBy do not exist in the register - the concepts are
  Obsoletes and IsObsoletedBy. Every emitted IRI and label was rechecked against the register.

Fields that were being dropped, now read: the contact person and the repository contact, the
owning collection, dc.identifier.other, dc.source.uri, local.demo.uri, every dc.relation
qualifier, dc.format, dc.coverage.placeName, given/family name parts, language names,
local.sponsor, local.additional.metadata, local.size.info, local.refbox.format and the metashare
tree. Bitstreams of the ORIGINAL bundle now become distributions with format, size and checksum.

Portability, none of which changed LINDAT output: ISO 639-1 and region-tagged language codes are
accepted, the stock DSpace type list is mapped, dates are proved castable before being emitted,
the funding guard names the whole grantAgreement prefix, and a repository with no url is no longer
described as living at each item's own address.

Tests: 70 methods over 9 fixtures, including a stock-DSpace record, every date shape DSpace
stores, and the two CLARIN access shapes. Element order is asserted directly - the CCMM schema is
an xs:sequence and validating against it would pull xml.xsd and GML over the network at build
time - plus closed-set guards so no off-register string can come back. Mutation testing: 45 of 46
single-edit mutants are now caught, against 20 of 72 before.

Re-run over all 2628 live records: 2628/2628 valid against CCMM 1.1.0 in XSD 1.1.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

XmlMatcherBuilder builds its NamespaceContext from a plain map with no built-in prefixes, so
JAXP cannot resolve the reserved "xml" prefix and ccmmXmlLangFollowsTheXoaiLanguageWrapper
failed with "Prefix must resolve to a namespace: xml". Register it alongside ccmm.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

…tested

Copilot asked for coverage of the embargo/restricted access paths and of the FormatDate
variations. Most of what it listed is already asserted, and one of its premises was wrong -
the crosswalk never reads local.embargo.termslift, embargo comes from others/access-status.
Three branches genuinely had no assertion, so they get one:

- others/restrictedAccess=true on a PUB record with files must still be restricted access;
  the download gate outranks the licence category. 19 live records take that path.
- others/access-status=embargo must reach the COAR embargo term, not the restricted fallback.
- a dc.date value carrying a time must stay a ccmm:date_time; time_instant is a choice, and
  truncating it into ccmm:date would silently drop the time.

73 test methods over 11 fixtures.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

license-maven-plugin checks src/** and fails the dspace-oai build without it. The check runs in
the verify phase, so it only surfaced once the unit tests stopped failing first.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Every record already validated against dataset/schema.xsd, which is why none of this showed up
as invalid. The CCMM specification states requirements the XSD has no way to express - they live
in en/dsv.ttl as scope notes and cardinality - and a conformance check reads those, not the XSD.
All 89 scope notes and 92 cardinality statements were extracted and checked against the output of
all 2628 live LINDAT records.

The one the specification calls mandatory:

- FRASCATI FORD. Dataset.hasSubject: "At least one subject must be a value from FRASCATI FORD
  vocabulary. For that subject use reference codelist .../SubjectCategory/ as the Subject scheme."
  No record carried one. The field is read off dc.type: a tool or service is computer science
  (10201), a newsreel clip is History (60101) - on this corpus every dc.type "clip" is a segment
  of a digitised 1943 newsreel held as a historical source, not a language resource - and
  everything else is Linguistics (60203), which is what the repository collects. Codes and both
  prefLabels come from the live register, and FORD_DEFAULT_CODE is the single value another
  deployment changes. 1161 History, 1136 Linguistics, 331 Computer sciences.

Statements the output was making that the source does not support:

- access_url pointed at the bitstream. accessUrl: "The resource at the access URL contains
  information about how to get the Dataset. It shall be the web page (not a document file)", and
  the schema's own Czech documentation says to use it "in all cases except a downloadable file
  URL". It is now the item's landing page; download_url keeps the file.
- identifier/value carried the resolved URL. CCMM types value as skos:notation - the identifier
  WITHIN its scheme - and keeps the resolvable form in iri, the split its own samples use. One
  emitter now serves every source, so scheme, notation and IRI cannot disagree between branches.
- The dataset had no iri of its own, so the published record had no resolvable identity.
- license is 1..1 and was serialised as <license/> on 584 records. A void mandatory element tells
  a harvester nothing; those records now carry rightsstatements.org UND, which is the standard
  statement for rights not stated.
- Language labels were DSpace's own strings, not the register's, and glued words together
  ("UpperSorbian", "AncientGreek (to 1453)") on 537 elements; their positional pairing could also
  slip once duplicate codes were removed. The scope note asks for the register's "label and/or
  its 3-letter code" and CCMM's own dataset-mini sample writes primary_language with an iri and
  no label, so the IRI alone is published.
- media_type published IANA IRIs for six types RFC 6838 reserves as unregistered (application/
  x-gzip and friends); those IRIs 404. They are suppressed, and the mandatory EU format element
  carries the file type - which is the scope note's own "otherwise Format MAY be used" branch.
- format carried the IANA media type where Format.label asks for a label from the EU file-type
  codelist. format is now the EU concept, media_type stays IANA. All 9125 distributions map to a
  real EU concept; every code was verified by content, because that register answers 200 for any
  path.

Values that were not what they claimed to be:

- 203 byte-identical sibling elements across 161 records, because DSpace stores one value under
  several language wrappers and two sources can carry the same value. Deduplication now happens
  where each block is assembled, so it also catches the cross-source repeats.
- "Germany||United States" was published as one place name; it is two.
- 20 alternate titles said nothing about how they differ from the title.
- dc.relation.uri was listed as a qualifier but had no term, so those related resources came out
  untyped; the register has Other for exactly that.
- An iri only had to start with a URL, so "http://a/ Book and enhanced publication" was published
  as an identifier. The whole value now has to be one URI.
- A packed contact field put an affiliation in the email slot.
- A second dc.title never became an alternate title. The predicate was
  doc:field[@name='value'][position() > 1], but position() counts within each parent and every
  XOAI language wrapper holds exactly one field, so it was always 1. Two titles only worked when
  both sat in the SAME wrapper, which XOAI does not do for values differing in language. Latent
  on this corpus - no record has two dc.title - but wrong, and the new fixture exercises it.

Measured after the change over all 2628 records: 2628 valid, every counter zero, 95 tests and
187 assertions green, 67 of 68 stylesheet mutants killed, 23 portability inputs clean. The
surviving mutant edits a FormatDate branch its callers make unreachable by pre-filtering.

Not changed, and why: eight records carry U+FFFD in title and description text. That is in the
source data - eight records in, eight records out - so guessing the intended character would be
inventing content. It belongs in the repository's metadata.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>