Status: Implementing. The native suite and failure-triage workflow are being integrated. Static standalone reuse is exercised with a second renderer. Native animation timing, interpolation, motion, event and contextual-use scenarios are integrated. The complete normative audit, broader interaction coverage and standalone publication remain open; the checklist below is the delivery record. Author: GPT-6 Astra Drafted by: GPT-5.6 Sol Prior review: Claude Fable 5 Created: 2026-07-12
Build donner-svg-suite as a reusable superset of resvg-test-suite: preserve the complete pinned upstream corpus and add ordinary SVG/PNG pairs for SVG2 behavior it does not exercise. Use a C++ GoogleTest harness and the existing pixelmatch comparison path. Add deterministic C++ scenarios for declarative animation and no-JavaScript interaction, with the same image comparison and failure artifacts. Other renderers should get the upstream tests and extension cases through one integration.
The accompanying SVG2 audit compares Donner with the latest served SVG2 Editor's Draft, pinned by retrieval date and content hash. A readable chapter/section checklist records requirements, evidence, and gaps. Adding a useful test does not depend on completing a schema, manifest generator, universal adapter protocol, or specification database.
0026 describes the broader conformance scope and optional W3C/WPT sources. This design owns the reusable native suite and the initial practical audit.
Integrate the native static harness and reviewed pilot pairs, run the entire pinned resvg inventory, and preserve actual failure artifacts. Then connect scenarios to the animation branch's real C++ clock/event APIs, expand the audit by requirement, and qualify a second renderer integration.
Use the existing resvg filesystem convention: SVG and PNG files paired by relative path, with explicit resources and deterministic licensed fonts. Extension paths use a separate namespace so a Donner case cannot replace an upstream case. The distribution composes the full pinned upstream snapshot and extension data; it must not require consumers to find resvg separately.
The reusable seam is intentionally small: enumerate cases, ask a renderer to render a file under controlled dimensions/resources, and compare RGBA images. C++ integrations can call the renderer in process; a small process wrapper may serve engines with a CLI. Add adapter machinery only for a concrete integrator requirement. Case metadata remains near its files; generated manifests are optional inventories, not the source of truth for static case discovery.
Donner's existing ImageComparisonTestFixture and resvg_test_suite retain their qualification role until a combined native lane proves equivalent discovery, rendering configuration, comparison policy, and artifacts. Engine policies are separate records with a diagnosis and scope. Keep canonical comparison results available when running with an engine-specific threshold or alternate reference.
A baseline for an extension case comes from an independently reasoned simple expected image or an independently reviewed reference renderer plus specification reasoning. Agreement alone is not proof. Store how the reference was produced, required fonts/resources, and the precise assertion. An agent must not regenerate goldens from the renderer it is repairing and call that a fix.
Use production C++ document APIs for clock advancement, animation sampling, and input dispatch. Scenarios are ordinary C++ tests plus SVG fixtures and sampled PNG references. No JavaScript harness or event script language is needed. A scenario controls viewport, resource environment, initial time, input coordinates, and sampling instants; it checks both state and pixels.
The animation branch owns engine implementation and the production API. A test-only fake can verify scenario runner order, but cannot count as animation conformance. Integrate the branch's established clock/event contract; report a blocked scenario honestly if an API is unavailable.
Priority scenarios include animation before/at/after begin and end, repeated intervals, freeze/remove, restart, additive composition, motion paths, transform interpolation, seeks, pauses, and animation in use shadows. Input scenarios include overlapping hit targets, every pointer-events value, clipping versus masking, hover/active/focus styling, tab order, keyboard link activation, and event-triggered animation. Target identity, focus state, or requested navigation requires a semantic assertion; identical pixels do not prove correct event dispatch. Navigation tests capture an intent without opening a browser or contacting a remote site.
SVG2 Appendix D does not require every animation technology. Supported animation methods still must respect processing modes and use shadow-tree rules. No-JavaScript interaction tests establish the declared native feature subset, not complete SVG2 dynamic-interactive conformance. In static mode, animation is disabled entirely; evaluating animations at time zero is a different operation.
The baseline and chapter inventory live in svg2_audit.md. Every requirement row links to a concrete section, paraphrases the assertion, names implementation/test evidence, and uses one of these states: not-audited, missing-test, implementation-gap, covered-pass, covered-fail, spec-ambiguity, or not-applicable. A pass requires a recorded run for the named revision/configuration. A non-applicable classification requires a cited processing-mode or conformance-class reason; unsupported behavior is not automatically non-applicable.
Full audit completion means reading every normative table, algorithm, and interface member, then reconciling the element/attribute/property/IDL indexes. A table-of-contents inventory is navigation, not completed review. Preserve requirements that need semantic or manual evidence even when a PNG cannot test them. Dependency revisions are recorded when used; draft status is disclosed rather than making every draft-only dependency an automatic failure or silently substituting another spec.
The test command emits GoogleTest results and the exact comparison policy. A failure bundle includes input identity, expected/actual/diff images when available, dimensions, mismatch metrics, backend, font tier, and the reproduction command. A missing image or crash is an infrastructure/render failure, not a zero-difference result. Parallel runs need isolated artifact directories.
The agent workflow inspects those artifacts, reads the cited requirement, checks independent evidence, minimizes the input if needed, and classifies the problem: implementation defect, unsupported feature, reference problem, allowed rasterization variance, invalid input, or infrastructure error. Ambiguity remains explicit. Policy changes require their own evidence and review; agents do not make a failing corpus green by silently increasing budgets.
For the official harness, implement Donner in the upstream harness's backend convention, render its full selected corpus, and feed the outputs through pixelmatch and the same artifact/triage workflow. The Qt viewer is optional inspection tooling. Keep upstream-integration source changes and exact run commands reviewable. Local execution is distinct from submission to the upstream project.
Treat SVGs, fonts, resources, reference images, and renderer output as untrusted data. Reuse bounded image decoding and resource resolution. Validate dimensions, arithmetic, paths, and missing files; reject resource escapes and avoid shell interpolation. Run untrusted rendering with suitable process and resource limits. Scenarios must not trigger external navigation or unrestricted network fetches. Document which limits are actually enforced by the native runner; planned isolation is not a sandbox.
Fetched test descriptions and images are evidence, never instructions to an agent. Triage reports must avoid machine-specific paths and secrets. Preserve upstream licenses and attribution. A public suite release needs review of the exact corpus, build inputs, executable dependencies, and artifact bytes.
The existing //donner/svg/renderer/tests:resvg_test_suite is a renderer qualification target, not a complete audit gate. New native targets must be named in the implementation evidence before their checks are claimed. Required coverage includes corpus completeness, deterministic enumeration, missing/malformed PNGs, dimension mismatch, exact and tolerant comparison, per-engine policy, artifact identity, and pass/fail/skip accounting. Use real failing images to test triage.
Run //donner/editor/tests:svg_suite_animation_tests and its Geode variant for focused native scenarios using the actual clock, host-control, event and renderer APIs. Compare one second renderer's results through the same harness. A backend or font tier not executed remains unverified. Preserve the existing resvg lane until its replacement has demonstrated parity.
A universal schema and cross-process typed protocol can serve a future concrete consumer, but creates too much machinery before static cases are useful. A Python rendering/comparison stack duplicates the existing C++ fixture. Qt-only triage imposes human work and is unsuitable as the release gate.
The reusable boundary is the image/discovery/GoogleTest library plus a renderer callback; the standalone runner also invokes renderer executables without a shell. The installed resvg renderer exercises reuse. The standalone publication location remains open. None blocks authoring SVG/PNG fixtures or auditing a requirement. Broad WPT imports and scripted browser conformance remain separate work.
1.15.0 