Native interoperability validation

pnpm test:native runs the independent native suite in tests/native/. CI uses Node 24 separately from core Node 20/22/24 checks. Use Node 24.15+ locally for the current DOM validator. Standard pnpm test excludes native tests; repository TypeScript checks include their source. Validators remain dev-only dependencies; core and published format modules do not import them.

pnpm test:coverage runs standard and native tests together on Node 24 and checks converter coverage against committed thresholds. Core Node 20 tests remain separate and exclude native modules. The initial native-foundation corpus measured 96.30% lines, 92.55% statements, 95.29% functions and 85.90% branches. The latest reviewed expanded corpus measured 97.10% lines, 92.26% statements, 95.23% functions and 86.44% branches across 4,222 tests on Node 24.15.0. These are distinct snapshots.

Global floors remain 95% lines, 91% statements, 93% functions and 84% branches. The committed configuration also sets line/branch floors for selected D2, Mermaid, XML, ELK and primary CSV/JSON adapters, plus adjacency matrix/core, Matrix Market, Graphviz JSON, draw.io, Pajek, PlantUML and all six expanded Mermaid profiles with their shared parser helpers, plus scoped GraphSON, DGML, CX2 and RDF core/syntax modules. Coverage output is ignored. Native semantic assertions complement coverage; neither proves full language coverage.

Python oracles use an isolated task environment, with exact versions in requirements/native.txt. CI installs it before combined coverage; missing or wrong versions fail explicitly. Local setup uses Python 3.12+:

python3 -m venv .venv-native
.venv-native/bin/python -m pip install -r requirements/native.txt
pnpm test:native

GRAPH_NATIVE_PYTHON can select another task-owned environment. No Python package is needed by the library at runtime. CSV/TSV native checks assert exact cell contents and the separate Graph projection; CSV parsing alone does not establish graph-domain validity.

Draw.io fixtures exercise maxgraph 0.25.0's native legacy mxGraph model import, including UserObject values, style records, grouping, routing and page inflation. This proves import compatibility for asserted fields, rather than rendering in the draw.io application. The unchanged draw.io generated-file XSD rejects the compressed-page fixture; the suite records that negative result separately from successful native inflation/import. Pinned Python igraph fixtures inspect Pajek vertices, sections and attributes through the native reader. These oracles assert model import and selected native fields; they do not establish complete draw.io rendering or unrestricted Pajek grammar. Graphviz fixtures inspect actual Graphviz JSON output; Graphviz has no native JSON input reader, so fresh structural JSON export has no such reader proof. PlantUML fixtures inspect actual compiler SVG and class XMI rather than accepting syntax alone.

Installed Graphology, Graphlib, Cytoscape, D3 and ELK provide independent runtime oracles. Their fixture tests record precise installed versions. XML/JSON schema fixtures include their own official source and provenance. Invalid legacy output profiles are tested explicitly; passing a negative validation control does not promote that profile to standards compliance.

Mermaid fixtures are independently authored from official syntax documentation, with URLs and oracle versions in mermaid-profiles.json. All eight base diagram dialects validate both native input and emitted output. Database checks assert selected graph labels, shapes, edge strokes, participant/message/note order, namespaces/members/multiplicities, ER attributes/cardinalities, state hierarchy/transitions, scoped block columns and tree structure. Exact Unicode, quotes and backslash counts use the pinned native entity decoder plus browser HTML decoding. This integration deliberately checks the native decoder's bundle signature when locating it; version upgrades require review. It is test-only. The corpus includes 251-node flow and 50-level tree profiles, adjacent quoted block labels ending in literal backslashes, malformed controls and edited document output. Native flow comments occupy separate lines; inline %% after a declaration is rejected by the pinned parser.

D2 fixtures record official documentation and compiler version in d2-profiles.json. Native compilation checks escaped labels, literal dotted identity, isolated nodes, SQL column types/constraints and connection column indices, mixed arrows/full arrowheads, scoped ancestor endpoints and variable/ class substitution. Additional profiles contain a 60-node chain and 20-level scope. The D2 wrapper exposes no public worker shutdown API, so compilation runs in a subprocess with a 25-second deadline and explicit exit after output flush. No private worker termination API is used.

Native acceptance and selected semantic assertions do not prove unrestricted Graph preservation or interpretation of all native syntax. Exact source-document retention is a separate contract.

Sources: Mermaid usage, D2 JavaScript wrapper, and the per-fixture official documentation links.

Six further Mermaid profiles cover Architecture, Requirement, Sankey, GitGraph, C4 and UseCase. Native database snapshots are captured before reparse so shared native database reset cannot create a false equivalence. Their assertions include semantic modifiers, literal text, graph hierarchy, canonical direction, native configuration and negative grammar controls. Mermaid12.1 Sankey has an ASCII-only lexer and trims even quoted outer whitespace; Unicode names reject projection and remain exact in GraphDocument source. Requirement empty text uses omission; the native parser rejects an explicit text: "" field.

PlantUML is a checksum-pinned development JAR at .tools-native/plantuml.jar. scripts/native/plantuml-tools.json records primary release URLs, hashes and licenses. CI uses Temurin Java21; local macOS ARM64 setup can use the separately checksum-pinned task JRE. No system Java installation or hosted renderer is needed:

pnpm setup:native:plantuml # existing JAVA_HOME / CI Java21
node scripts/native/setup-plantuml.mjs --jre # task-local macOS ARM64 JRE

The JAR/JRE, GPL development tooling and generated compiler outputs stay in ignored .tools-native/; none enter library runtime imports or the published package.

String-domain controls distinguish Unicode scalars from opaque JSON strings. Mermaid's native/DOM entity path replaces NUL and surrogate references; display label emission/projection/patches reject them, while astral characters and safe ID aliases remain supported. D2's compiler retains NUL label fields but replaces lone surrogates; its adapters reject only that proven surrogate loss. These tests inspect native compiled/display fields rather than asserting file-byte encoding.

GraphSON native setup uses pnpm setup:native:graphson: twelve Maven Central JARs pinned individually by SHA256, derived from the verified official Apache TinkerPop 3.8.2 distribution. The manifest records URLs, versions, distribution provenance and licensing. Tools stay in ignored .tools-native/ and never enter runtime imports or package contents. Java subprocesses have a 20-second bound; GRAPH_NATIVE_JAVA selects the task Java 21 executable. Actual Java mapper tests use TinkerIoRegistryV3 plus extended value types, preserve unsafe magnitudes in raw serialized strings, and assert observed native topology/property semantics. No GraphSON adjacency/writeGraph interoperability is claimed.

DGML uses the official Microsoft 2009 XSD with xmllint, including negative schema controls for dynamic property attributes; this does not establish Windows renderer behavior. CX2 uses the maintained NDEx Python reader with observed default/null limitations. RDF uses N3 strict grammars and independent RDFLib parser/store observations and dataset isomorphism; empty-graph and language-case distinctions are checked before the native store normalizes them. Per-format READMEs and fixture provenance describe each bounded interpretation claim. Exact document source retention is separate from these native semantic assertions.

Final combined evidence belongs in the domain formats verification report. Isolated measurements do not establish a combined release gate; existing global coverage thresholds remain unchanged.