Golden artifact conformance

Renderflow's versioned golden corpus binds its artifact claims to a small, redistribution-safe acceptance harness. The corpus contains synthetic payloads only. It does not contain publication, comic, customer, or third-party content, and no fixture requires network or AI access.

For a small, retained source-to-output run through the public CLI and real local providers, see the inspectable artifact gallery. The gallery is explicitly invoked on a local machine; it complements this versioned conformance corpus and does not silently bless changed outputs.

Contracts

The canonical manifest is tests/fixtures/golden-artifacts/v1/corpus.json. Binary payloads are stored as lowercase hexadecimal inside the manifest so review, reproduction, and source control remain transparent. The manifest conforms to schemas/renderflow-golden-corpus-v1.schema.json.

The focused Rust harness materializes those payloads in a temporary directory and exercises production intake, graph planning, execution, validation, cache, checkpoint, hygiene, profile, and Flow-projection APIs. It can emit a report conforming to schemas/renderflow-conformance-report-v1.schema.json:

RENDERFLOW_CONFORMANCE_TIER="fast" \
RENDERFLOW_CONFORMANCE_REPORT="/tmp/renderflow-conformance.json" \
cargo test --package renderflow --test golden_conformance --locked

The report pins the corpus digest and engine version. Each materialized fixture records its artifact identity, payload digest, detected format, media type, and expected outcome. Each scenario records its status, assertions, and selected providers. Tool-backed runs include observed provider version strings; missing optional providers are unavailable, never silently successful.

Tiers

Tier	Purpose	Dependency policy
fast	Pull-request acceptance	Hermetic; Rust toolchain only
tool_backed	Adapter verification	Optional tools are probed and versioned
maximal	Scheduled and release evidence	All corpus families and scenario states are reported

The dedicated conformance workflow runs the fast tier for pull requests and the maximal tier on its schedule, on published releases, and when manually dispatched. Both publish the machine-readable report even when a job fails.

Adding a fixture

1. Use an original synthetic payload small enough to inspect in review.
2. Add its payload, family, expected identity, and outcome to the versioned corpus manifest. Never add secrets, protected references, or production content merely to exercise a gate; use an unmistakably synthetic marker.
3. Map at least one scenario to the fixture and add an assertion through a public production API.
4. Validate both schemas and run the focused harness locally.
5. Increment corpus_version when fixture bytes or expected behavior changes.

A capability graduates from unavailable or experimental only when the corpus proves its success path, failure classification, validation evidence, and provider behavior on every supported platform. Nondeterministic formats must assert structural and semantic invariants rather than byte identity.