native

Native runner profiles (experimental)

These adapters construct native invocations and decode bounded results. They do not admit executables, capture source/build dependency closure, or grant CEM/OCM semantic authority. The caller must bind the executable and independently prepared project/binary inputs, provide a fresh report directory and enforce containment.

ctest executes the configured build directory, not a build step. googletest and catch2 execute independently built test binaries. Their child binaries, libraries, configuration, fixtures and environments remain caller bindings. dotnet-vstest-* uses the VSTest profile with --no-build --no-restore; it does not qualify Microsoft.Testing.Platform or arbitrary browser drivers. TRX results join native test IDs to framework-specific definitions. Parameterized rows sharing the same method identity abstain rather than silently collapse.

cargo-test is a single library target; cargo-doctest is its separate Rustdoc suite. cargo-integration and cargo-bin require the literal Target name and execute one integration or binary test harness. Stable libtest pretty output is parsed with exact terminal counts and serial execution, with incomplete build/collection output retained as incomplete. Custom harness output and multi-suite invocations are unsupported. Stable Cargo JSON describes compiler artifacts, not test outcomes: Rust 1.98.1 rejects libtest JSON without a nightly compiler. No nightly feature is enabled silently. nextest writes a fixed profile-owned TOML under the fresh report directory, disables retries for execution and includes skipped tests in JUnit. Its parser also preserves reported native retry extensions. It does not run doctests.

Go observes go test -json -count=1 for one package. The existing independently qualified Go live-test provider remains a separate authority path. Package/test identities, incomplete events and build failures remain explicit. Subtest filter segments require a separately specified selection profile and are rejected here.

Retained live format evidence

The testdata reports were generated on macOS arm64 on 2026-10-01 from disposable pass/fail/skip fixtures (GoogleTest/CTest also disabled), using:

• Go 1.27.1: go.json.
• CTest 4.4.3: ctest.xml.
• GoogleTest 1.18.0: googletest.xml.
• Catch2 3.16.0: catch2.xml.
• .NET SDK 9.0.316 / Test SDK 17.14.1: NUnit 4.3.2 with adapter 5.0.0, MSTest 3.6.4 (nunit.trx, mstest.trx).
• SDK template xUnit 2.9.2 with its VSTest adapter (xunit.trx).
• Cargo/Rust 1.98.1: cargo.txt, doctest.txt.
• nextest 0.9.146: nextest.xml, plus an always-failing retry fixture nextest-retry.xml with two native failed attempts.

These prove concrete native formats, not complete runner qualification. Native runtime coverage for zero tests, build/collection failures, parameterization, flaky recovery, process timeout/interruption/descendant retirement and every framework/platform version remains incomplete. Unit negatives retain those gaps; they are not substituted for live lifecycle qualification. Parent coordinator lifecycle evidence must remain separately identified.

Official contracts: GoogleTest docs/advanced.md; Catch2 docs/reporters.md and docs/ci-and-misc.md; CMake Help/manual/ctest.1.rst; Microsoft .NET docs/core/tools/dotnet-test-vstest.md; nextest machine-readable JUnit/list docs. The source frameworks were fetched from their official tagged repositories; Rust/nextest executable downloads were verified against official SHA256 files.

Requirement evidence

• TRE-V0-001/003/004/005: TestBuildExactSelectorsAndFixedProfiles and TestQualifiedIdentitySelectors exercise fixed argv/configuration and exact native identity. Caller file creation/executable admission remains common code.
• TRE-V0-006/007: TestLiveNativeReports, TestNextestNativeRetriesRetained, TestRejectsIncompleteAndContradictoryReports, TestGoStructuredStreamAndBuildFailure and TestCargoIncompleteAndCountMismatch retain status/count/identity failures.
• TRE-V0-002/008 are coordinator-owned; this package claims no source-closure or process-retirement proof. TRE-V0-009 is PARTIAL per the live matrix above.
• TRE-V0-010 is NOT_PRODUCED here: runner observations alone do not establish assertion mappings, causal criterion kills or Tasks completion authority.

Go selectors are exact package::TestName (the package must equal Project). Nextest selectors are exact binary-id::test-name; its native binary_id filter was exercised with nextest 0.9.146. Cargo selectors identify a test within the explicit single selected target. Report identity cannot imply across-target uniqueness without the caller's project/target/executable binding.

Independent review repair evidence

Review cycle 1 corrected the shared args-only invocation contract and declared native failed-test exits (CTest 8, Catch2 42, Cargo 101, nextest 100; other admitted profiles 1). TestNativeGoBuildExecuteParse now exercises Build → the shared executor → Parse → Normalize with real pinned Go/source bytes and pass/fail/skip results. This is a Go host execution witness, not qualification of other tuples. Go uses the local toolchain with module-proxy and sum-database network access off.

The TestReview* regressions cover native GoogleTest result/status/skip/disabled contradictions, TRX summary and counter conflicts, nextest crash retry causality, Cargo announced inventory, duplicate Go structural keys and package terminals, and profile-specific exit codes. Retry attempt messages retain nextest's native type and text; an ordinary nonzero process exit remains UNKNOWN cause, while signal/abort/setup/timeout evidence is INFRASTRUCTURE, never inferred ASSERTION. The additional nunit-pass.trx was emitted by the same NUnit runtime using an exact passing method filter; its native successful summary is Completed.

CMocka 2.0.2 (experimental dedicated tuple)

cmocka-xml runs a separately built and pinned regular native executable in one explicit primary TEST phase with no argv. It fixes CMOCKA_MESSAGE_OUTPUT to STANDARD,XML, CMOCKA_ERROR_OUTPUT to STDERR and a fresh cmocka.xml path. The existing Target field binds one group; exact native group::test identities and 1..64 expected tests are mandatory. One literal selection is supported; wildcards, multiple selections and project/config/reporter/tool overrides are not.

Both original STANDARD streams and native XML must agree. Unrecognized output is incomplete for this bounded profile. Native XML failure causes remain UNKNOWN; fixture errors, including group teardown returning nonzero with native exit 0, cannot become passing observations. Shared Normalize turns incomplete public states UNKNOWN without rewriting the original reports. No retry is scheduled or inferred. Per-test source paths are not invented from C function names.

Nine actual macOS arm64 cases retain mixed outcomes, exact selected pass/skip, zero tests, per-test setup/teardown, group setup/teardown and wrong-target evidence. The native source is CMocka 2.0.2 immutable commit fefa2b8a023121f7235e18ed17249e4012dd144f; all 150 official release files matched that official commit archive. Apache-2.0 and upstream CMake redistribution notices were retained. Publisher PGP signatures were acquired but not authenticated; the owner accepted this limit for the bounded trusted-local experiment only. Static-build provenance is separate from TEST receipts; SDK/loader closure and new timeout/interruption qualification remain unobserved. See the CMocka provenance fixture and TRE-V0-015..017. Unity and complete C-family coverage are not delivered by this profile.

Ginkgo v2 (experimental, unqualified)

ginkgo-v2 runs one separately built and pinned Ginkgo v2 suite binary (a compiled Go test package) with 17 fixed argv elements and a fresh ginkgo.json JSON report. Target is GoWrapper::SuiteDescription; expected identities are Target::FullText, the space-joined container and leaf texts. Selection adds one -ginkgo.focus=^<QuoteMeta(SuiteDescription)> (<QuoteMeta(FullText)>|...)$ element, bounded at 4096 bytes. Build requires a pinned literal _test.go Project, an absolute symlink-resolved Root and fixes offline Go module settings; configuration, reporter, report-file and tool overrides are refused. Empty argv is never emitted.

The parser accepts exactly one report holding one suite, decodes strict JSON with exact enum tables (null refuses), and checks SuiteDescription, SuitePath, the full SuiteConfig, counts, inventory, eligibility, suite-level nodes, failure context and exit/suite coherence. Hook failures, retries, parallel processes, AdditionalFailures, programmatic focus, special suite failure reasons and eligible specs that never ran (including Ordered follow-on skips) are incomplete. AdditionalFailures are emitted only in verbose mode, so their absence is not proof. Failure causes remain UNKNOWN; no retry is scheduled.

Fixtures are synthetic and shaped from the Ginkgo v2.33.0 source reading pinned in testdata/ginkgo-provenance.json. Actual runner execution is NOT_RUN because no Ginkgo binary or module was available locally; the buildinfo module-sum check is NOT_IMPLEMENTED. See TRE-V0-018..020.