v1.29.0

[1.29.0] - 2026-08-30

Shared MCP listener for multi-client use, PACER document-mapping
fidelity (docket roster, identity dedup, account-handle hygiene),
and XSD-canonical hexBinary literals so SPARQL joins match across
extractors and builders.

Shared MCP listener

• The CASE/UCO MCP server can run as a shared SSE/HTTP listener
  (scripts/run-mcp-server.sh, default http://127.0.0.1:8765/sse)
  so Cursor, Hermes, Claude Desktop, and other MCP clients use one
  process. Do not set a Linux PATH on a Windows wsl.exe stdio
  wrapper — that made Cursor fail with spawn wsl.exe ENOENT.
  PATH and CASE_UCO_EXTENSIONS are set inside the Linux process.

PACER document mapping

• Raised process_document_file MAX_BYTES from 10 MiB to 50 MiB so
  scanned PACER plea agreements and translations are not refused as
  source_oversized.
• route_investigation_content now requires token boundaries for short
  or numeric family keywords (ttp, c2, 846, 554), so routing a
  Layer-1 JSON-LD graph no longer false-hits CTI or export-control from
  https:// IRIs and statute fragments.
• PACER semantic mapping now extracts multi-title USC/CFR citations,
  UNITED STATES v. captions, named forensic tools, and platform
  accounts beyond Snapchat. Shared output folders also get
  stem-specific extraction sidecars so one PDF does not overwrite
  another's extracted-content.json.
• Added extract_pacer_docket_roster, which parses the PACER criminal
  docket structurally: defendant number, name, also known as aliases,
  counsel, and the docket-entry-1 per-defendant count matrix. A PACER
  defendant block contains its own represented by attorney block, so a
  string-level reading promotes defense counsel to the charged party.
  Layer-2 synthesis should read defendants and counts from this parser
  rather than from the general NER pass. Firm and address lines are
  rejected as attorney names, and the roster is empty for non-docket text.
• Added jurisdiction, court, and venue tokens to PERSON_NAME_STOPWORDS
  so captions no longer yield uco-identity:Person nodes named
  "New York", "New Mexico", "Eastern District", or "District Court".
• Fixed a relationship @id concatenation bug in
  scripts/build_pacer_layer2_batch.py that produced ids of the form
  kb:rel-def-kb:charge-1. Affected docket graphs were rebuilt.
• Semantic mapping now resolves identity for repeated mentions. Minor
  victim labels, PACER case numbers, Document N Filed ... stamps, and
  date references each yield one node anchored at their first mention.
  Previously an indictment emitted a separate @id for every mention, so
  "Minor Victim 4" became nine victims and one page header became twelve
  events. Because extraction is capped at MAX_SEMANTIC_ENTITIES (96),
  that repetition also spent the budget on page furniture and dropped real
  entities appearing later in the document: on the ENTERPRISE indictment
  deduplication recovered Minor Victims 8, 9, and 10, which the capped run
  had truncated.
• PLATFORM_ACCOUNT_RE no longer turns prose into accounts. The handle
  segment was optional-prefixed and unvalidated, so narrative text yielded
  ApplicationAccount nodes named "Discord was", "Discord or", and
  "Discord calls". A token is now accepted only when the source marks it
  as an account (@ prefix, surrounding quotes, or a preceding
  "account"/"username"/"handle" keyword) or it carries handle-like shape,
  and never when it is a common English word. Handles no longer absorb
  sentence-final punctuation.
• Fixed scripts/_pacer_attach_wechat_source.py, which passed a Path to
  CASEGraph.load (expects a JSON string) and collided on the kb:
  prefix because the copied exemplar declares its own base.

Canonical hexBinary

• xsd:hexBinary literals are now emitted in XSD canonical (uppercase)
  form by CASEGraph, exposed as case_uco.graph.canonical_hex_binary.
  SPARQL joins compare literals as terms, so a digest written by
  hashlib.hexdigest() (lowercase) does not join to the same digest
  written by the document extractor (uppercase). The two layers therefore
  failed to connect in a triplestore even though the bytes matched.
  This is easy to miss because backends disagree: rdflib silently
  canonicalizes the literal and the join appears to work, while Oxigraph
  does not and returns nothing. Across the PACER corpus the mismatch hid
  35 of the document-to-investigation links; term-equality and
  case-insensitive joins now both return 248. Values that are not valid
  hexBinary pass through untouched, so Bitcoin transaction identifiers on
  cryptoinv:CryptocurrencyTransactionFacet keep their lowercase form.
  The hand-rolled lit() helpers in examples/pacer/*/build_*.py build
  JSON-LD without CASEGraph, so they now share the same helper, and the
  affected exemplars were rebuilt.

Change proposals and CI

• Added change_proposals/cac-enterprise-hierarchy-cardinality, a CAC
  proposal to make hasHierarchy and hasLeadershipRelation optional on
  ChildExploitationEnterprise. A charged 18 U.S.C. § 2252A(g) enterprise
  need not plead internal structure, and requiring those properties
  pressures a modeller to fabricate hierarchy to pass validation.
• Refreshed examples/cti/darkwatchman_2021/build-manifest.json so the
  modeling_guidance hash matches docs/recipes/cyber-threat-intelligence.md.
  The v1.28.0 recipe edit left the sidecar stale, and CI Test Python
  failed on CRIT-C-PROVENANCE-MANIFEST-MISMATCH.

Package versions bumped to 1.29.0.

---

Verification

All release artifacts include:

• SHA256 checksums (SHA256SUMS.txt)
• SLSA build provenance attestations via actions/attest-build-provenance
• Python SBOM (sbom-python.cdx.json) in CycloneDX format

sha256sum -c SHA256SUMS.txt
gh attestation verify <artifact> --repo vulnmaster/CASE-UCO-SDK

For full release notes including prior versions, see CHANGELOG.md.