Status: Draft Author: Claude Opus 4.6 (with SpecBot, PerfBot, BazelBot) Created: 2026-04-12

Summary

Comprehensive audit of the resvg test suite, rendering pipeline, and build infrastructure to identify actionable GitHub issues. This doc catalogs **~307 skipped tests** (plus ~51 render-only) across the resvg suite, distills them into 24 SVG feature/bug issues, and proposes 14 CI/build optimization issues for reducing CI runtime and binary size.

Key findings:

P0 quick wins (2 issues) would unblock 50 tests with small effort — CSS filter functions and transform-origin are already parsed but not applied at render time.
P0+P1 combined (7 issues) would unblock **~139 tests** — roughly 60% of all failures.
**~30 tests** are intentionally skipped (deprecated SVG 1.1 features) and should stay skipped.
CI wall-clock time is dominated by macOS runner queue wait (55–130 min); actual build+test work is only ~12 min (Linux) / ~9 min (macOS).
Compilation speed is bottlenecked by the monolithic entt/entt.hpp header (95,903 lines) included transitively by 48 source files.

Goals

Provide a prioritized backlog of issues that can be filed on GitHub.
Each issue has clear scope, test coverage impact, spec references, and complexity estimate.
Enable data-driven prioritization of v0.5+ work.
Identify quick wins that unblock the most resvg tests with the least effort.

Non-Goals

Actually filing the issues (this doc is the proposal for review).
Implementing any of the proposed changes.
Covering Geode backend gaps (tracked separately in the Geode design doc).
Animation or scripting support (out of scope for current milestone).

Next Steps

Review this doc and decide which issues to file.
File P0 issues first — they're small, high-impact, and have no dependencies.
File P1 issues and assign to upcoming sprints.

Part 1: SVG Feature Gaps & Bugs

Source data: donner/svg/renderer/tests/resvg_test_suite.cc (1,539 lines), 307 Params::Skip() entries, 51 Params::RenderOnly() entries, 9 elevated-threshold entries.

Intentionally Skipped — Do Not File (30 tests)

These are deprecated SVG 1.1 features. They should remain Params::Skip() permanently.

Feature	Tests	Spec Status
enable-background	17	Removed in SVG2 §11.7.7
<tref>	9	Removed in SVG2 §11.5
kerning attribute	2	Deprecated; replaced by font-kerning
glyph-orientation-*	2	Deprecated; replaced by text-orientation

P0 — Quick Wins (50 tests, small complexity)

These features are already parsed — the rendering path just needs to be connected.

Issue S1: Apply CSS filter function shorthands at render time

Field	Value
Tests unblocked	30 (filters/filter-functions/*)
Complexity	Small — this is a bug, not a new feature
Spec	Filter Effects §8 — Filter Functions
Already tracked	B2 in docs/design_docs/0021-resvg_feature_gaps.md

Current state: PropertyRegistry::ParseFilterFunction() correctly parses all 10 CSS filter functions (blur(), brightness(), contrast(), drop-shadow(), grayscale(), hue-rotate(), invert(), opacity(), saturate(), sepia()). The rendering path only recognizes ElementReference variants (SVG url(#id) form) and ignores the CSS function vector.

Key files: RenderingContext.cc:244, RendererDriver.cc Dependencies: None

Issue S2: Support transform-origin as a presentation attribute

Field	Value
Tests unblocked	20 (structure/transform-origin/*)
Complexity	Small–Medium
Spec	CSS Transforms 2 §6, SVG2 §5.1
Already tracked	F2 in docs/design_docs/0021-resvg_feature_gaps.md

Current state: Works in style="" attribute. LayoutSystem.cc computes and applies it correctly. The presentation-attribute parsing path (<rect transform-origin="50% 50%">) does not recognize it. Single-keyword syntax (center) doesn't parse either.

Key files: PropertyRegistry.cc, TransformOrigin.h Dependencies: None

P1 — High Impact (89 tests, medium complexity)

Issue S3: Fix <image> element sizing and layout

Field	Value
Tests unblocked	18 (structure/image/*)
Complexity	Medium — multiple sub-bugs
Spec	SVG2 §5.8, CSS Images 3 §5

Embedded data-URL images render but at wrong size. preserveAspectRatio modes need investigation. 5 of 18 tests involve <use> referencing inline <svg> elements with width/height/viewBox combinations.

Partially blocked by: B1 (non-square viewBox + percent resolution) Key files: LayoutSystem.cc, SizedElementComponent.h

Issue S4: Implement <switch> element + systemLanguage

Field	Value
Tests unblocked	15 (structure/switch/* + structure/systemLanguage/*)
Complexity	Medium
Spec	SVG2 §5.9 — Conditional Processing

Not implemented — no SVGSwitchElement in ElementType.h. For a standalone renderer, the language should be configurable (default: en). systemLanguage uses BCP 47 prefix matching.

Dependencies: None Key files: New SVGSwitchElement.h/.cc, ElementType.h

Issue S5: Implement dominant-baseline full keyword set

Field	Value
Tests unblocked	14 (text/dominant-baseline/*)
Complexity	Small — adding 4 mapped keywords
Spec	CSS Inline 3 §4.2

9 of 13 keywords work. Missing: before-edge → text-top, after-edge → text-bottom, text-before-edge → text-top, text-after-edge → text-bottom (SVG 1.1 legacy mappings).

Key files: PropertyRegistry.cc:198-215, TextEngine.cc:311-325

Issue S6: Implement context-fill / context-stroke rendering

Field	Value
Tests unblocked	12 (painting/context/*)
Complexity	Medium–Large
Spec	SVG2 §13.3 — Context paint

Parsed (PaintServer::ContextFill/ContextStroke variants exist). The marker rendering path does not propagate context paint. Heavily used by icon libraries.

Key files: RenderingContext.cc, PaintServer.h

Issue S7: Implement alignment-baseline full keyword set

Field	Value
Tests unblocked	10 (text/alignment-baseline/*)
Complexity	Small–Medium
Spec	CSS Inline 3 §4.4

Missing baseline (the initial value per spec), no-change, reset-size, use-script.

Dependencies: Issue S5 (shared enum expansion) Key files: PropertyRegistry.cc:1522

P2 — Medium Impact (38 tests, medium complexity)

Issue S8: Implement paint-order property rendering

Field	Value
Tests unblocked	8 (painting/paint-order/*)
Complexity	Medium
Spec	SVG2 §13.11

Parsed as an unparsed CSS string, not applied at render time. Controls fill/stroke/markers paint order. Commonly used for "stroke behind fill" text effects.

Key files: All three renderers, PropertyRegistry.cc

Issue S9: Fix mask-type property + mask edge cases

Field	Value
Tests unblocked	6 (masking/mask/*)
Complexity	Medium
Spec	CSS Masking 1 §6.1

Property not defined in PropertyRegistry. Sub-issues: mask-type=alpha, mask-type in style, mask color-interpolation in linearRGB, mask-on-self, mask-units edge cases.

Key files: RendererDriver.cc (mask compositing), PropertyRegistry.cc

Issue S10: Fix feMorphology edge cases

Field	Value
Tests unblocked	5 (filters/feMorphology/*)
Complexity	Small
Spec	Filter Effects 1 §15

Core erode/dilate works. Failures on boundary conditions: empty radius, non-numeric radius, zero radius (should be passthrough), negative radius.

Key files: FilterSystem.cc, FilterGraphExecutor.cc:483-497

Issue S11: Fix <use> referencing inline <svg> edge cases

Field	Value
Tests unblocked	5 (structure/use/*)
Complexity	Medium
Spec	SVG2 §5.6

Shadow tree instantiation edge cases with <use> referencing inline <svg> elements with various width/height/viewBox combinations.

Dependencies: Issue S3 (shared viewport resolution logic) Key files: SVGUseElement.cc, ShadowTreeSystem.cc

Issue S12: Fix intrinsic sizing + percent resolution

Field	Value
Tests unblocked	5 (across 3 categories)
Complexity	Medium

For SVGs with non-square viewBox and no explicit width/height, Donner's intrinsic document size comes out wrong. Already tracked as B1 in 0021-resvg_feature_gaps.md.

Key files: LayoutSystem.cc, SizedElementComponent.h

Issue S13: Fix stroke-dasharray edge cases

Field	Value
Tests unblocked	4 (painting/stroke-dasharray/*)
Complexity	Small
Spec	SVG2 §13.5.6

Edge cases: odd-length arrays (must be doubled per spec), zero-length dashes, all-zero sums (treat as solid line).

Issue S14: Fix <filter> edge cases (filterRes, filterUnits, multiple inputs)

Field	Value
Tests unblocked	4 (filters/filter/*)
Complexity	Medium

Skipped tests: content-outside-the-canvas-2, in=BackgroundAlpha, with-mask-on-parent, with-transform-outside-of-canvas.

Issue S15: Fix text rendering edge cases (mixed inline, BiDi-adjacent)

Field	Value
Tests unblocked	4 (text/text/*)
Complexity	Medium

Filter bbox, ligatures in mixed fonts, percent values on dx/dy, real text height.

Issue S16: Implement clipPath with <text> children

Field	Value
Tests unblocked	4–7 (masking/clipPath/*)
Complexity	Medium–Large
Spec	CSS Masking 1 §7.2

ClipPath works for shapes but text children are not supported.

Dependencies: Text rendering pipeline

Issue S17: Fix feImage edge cases

Field	Value
Tests unblocked	3 (filters/feImage/*)
Complexity	Small–Medium

Subregion with rotation, x/y with protruding subregions. Also feImage with external href (separate from ResourceLoader story).

Issue S18: Implement <a> element rendering

Field	Value
Tests unblocked	3 (structure/a/*)
Complexity	Small — treat as <g> for rendering
Spec	SVG2 §5.4

Not implemented. For rendering purposes <a> just needs to be a grouping element.

P3 — Lower Impact / Higher Complexity

Issue S19: Implement textLength + lengthAdjust (text stretching)

Field	Value
Tests unblocked	4 (text/textLength/) + 3 (text/lengthAdjust/) = 7
Spec	SVG2 §11.6.2

Issue S20: Implement textPath SVG2 features

Field	Value
Tests unblocked	5+ (text/textPath/*)
Complexity	Medium–Large per sub-feature
Spec	SVG2 §11.7

Sub-features: method="stretch", spacing="auto", side="right", path attribute (inline path data), referencing <rect>/<circle>.

Issue S21: Implement BiDi text (direction + unicode-bidi)

Field	Value
Tests unblocked	5+ (across text/direction, text/unicode-bidi, RTL tests)
Complexity	Large — requires Unicode Bidirectional Algorithm
Spec	CSS Writing Modes 4 §2, UAX #9

Properties registered but ignored. Consider scoping as "basic LTR/RTL" first.

Issue S22: Implement text-decoration SVG2 decomposition

Field	Value
Tests unblocked	2 (text/text-decoration/*)
Spec	CSS Text Decoration 3

Basic underline/overline/line-through works. Independent color/style not supported.

Issue S23: Implement font-kerning property

Field	Value
Tests unblocked	2 (text/font-kerning/*)
Complexity	Small (for --config=text-full)

Would toggle HarfBuzz's kern feature.

Issue S24: Implement image-rendering full value set

Field	Value
Tests unblocked	2 (painting/image-rendering/*)
Complexity	Small
Spec	CSS Images 3 §5.3

Only pixelated → nearest-neighbor is implemented. Missing crisp-edges, smooth/auto.

SVG Feature Impact Summary

Priority	Issues	Tests Unblocked	Cumulative
P0	S1, S2	50	50
P1	S3–S7	69	119
P2	S8–S18	~51	~170
P3	S19–S24	~23	~193
Intentional skips	—	30	—

Recommended Implementation Order

The dependency graph suggests this phasing:

Phase 1 (P0 — immediate):
  S1 CSS filter functions ──→ +30 tests
  S2 transform-origin attr ──→ +20 tests
 
Phase 2 (P1 — text baselines):
  S5 dominant-baseline ──→ +14 tests
  S7 alignment-baseline ──→ +10 tests (depends on S5)
 
Phase 3 (P1 — elements):
  S4 <switch> + systemLanguage ──→ +15 tests
  S18 <a> element ──→ +3 tests (trivial, bundle with S4)
 
Phase 4 (P1 — rendering features):
  S6 context-fill/stroke ──→ +12 tests
  S8 paint-order ──→ +8 tests
 
Phase 5 (P1-P2 — image/viewport):
  S12 intrinsic sizing ──→ +5 tests (unblocks S3, S11)
  S3 <image> sizing ──→ +18 tests
  S11 <use> edge cases ──→ +5 tests
 
Phase 6 (P2 — small fixes):
  S10 feMorphology edges ──→ +5 tests
  S13 stroke-dasharray edges ──→ +4 tests
  S9 mask-type ──→ +6 tests
 
Phase 7+ (P3 — as prioritized):
  S14–S17, S19–S24

Part 2: CI Runtime & Binary Size Improvements

Source data: CI workflow YAMLs, .bazelrc, build_defs/rules.bzl, MODULE.bazel, measured CI run times from run 24295860066.

Current CI State

Job	Platform	Queue Wait	Actual Work	Total
Linux	ubuntu-24.04	~0s	~12 min	~12 min
Linux-Geode	ubuntu-24.04	~0s	~2 min	~2 min
macOS	macos-15	55–130 min	~9 min	65–140 min
CMake (Linux)	ubuntu-24.04	~0s	~10 min	~10 min
CMake (macOS)	macos-15	varies	~10 min	varies
Coverage	ubuntu-24.04	~0s	~30–45 min	~30–45 min
CodeQL	ubuntu-latest	~0s	~10–60 min	varies

End-to-end CI: 65–191 min, dominated by macOS runner queue wait.

P0 — Critical

Issue B1: Reduce macOS runner queue latency

Field	Value
Impact	−60 to −120 min per CI run
Effort	1–2 hours

The single largest CI bottleneck is macOS runner queue wait (55–130 min), not build/test time. Actual macOS work is only ~9 min.

Options:

Use macos-latest (macOS 14) if the runner pool is larger.
Run macOS CI only on main merges, use Linux as the primary PR gate.
Defer fuzzer tests (2.3 min, separate LLVM 21 toolchain) to nightly/main-only.
Consider self-hosted macOS runners.

Key files: .github/workflows/main.yml

P1 — High Impact

Issue B2: Enable Bazel remote cache for CI

Field	Value
Impact	−30–50% build time on cache-hit runs
Effort	2–4 hours

No remote cache is configured. Every CI run rebuilds from scratch on cache miss.

Options: GitHub Actions Cache via bazel-contrib/setup-bazel, self-hosted bazel-remote, or Google's free RBE tier for open-source.

Key flags: --remote_cache=<url> --remote_upload_local_results --remote_timeout=300 Key files: .bazelrc, .github/workflows/main.yml

Issue B3: Increase --local_test_jobs from 2 to HOST_CPUS

Field	Value
Impact	−20–40% test execution time on Linux
Effort	5 min (one-line change)

Linux CI artificially limits test parallelism to 2 despite 4-core runners. The test step is 7.8 min — the dominant cost of the Linux job. No tests use local = True or tags = ["exclusive"].

Key files: .github/workflows/main.yml:75,125

Issue B4: Switch from monolithic entt/entt.hpp to modular EnTT headers

Field	Value
Impact	−30–50% compile time for ECS-touching TUs
Effort	2–3 days

EcsRegistry.h includes the monolithic 95,903-line single-include header. 48 non-test files transitively include it. Donner only uses entt::entity, entt::basic_registry, entt::basic_handle. The modular headers (73 files) are already vendored.

Replace with:

#include <entt/entity/fwd.hpp>      // 290 lines
#include <entt/entity/registry.hpp>  // 1,239 lines
#include <entt/entity/handle.hpp>    // 431 lines

Also create EcsRegistry_fwd.h for headers that only need type names.

Key files: donner/base/EcsRegistry.h, donner/svg/AllSVGElements.h

Issue B5: Add --config=ci with CI-specific optimizations

Field	Value
Impact	Maintainability + tuning
Effort	30 min

CI-specific flags are scattered across workflow YAML. Centralizing in .bazelrc as --config=ci makes them testable locally.

build:ci --local_test_jobs=HOST_CPUS
build:ci --jobs=HOST_CPUS
build:ci --verbose_failures
build:ci --repository_cache=~/.cache/bazel-repo

Issue B6: Add test sharding to top 5 unsharded test suites

Field	Value
Impact	2–3× test speedup for largest suites
Effort	30 min (one-line change per target)

Only resvg_test_suite_impl has shard_count = 16. All other tests run single-threaded.

Target	Test Cases	Recommended Shards
base_tests	~445	10
svg_tests	~417	10
parser_tests	~176	5
core_tests	~140	4
css_tests	~104	3

GTest supports sharding natively via GTEST_SHARD_INDEX/GTEST_TOTAL_SHARDS which Bazel sets automatically.

P2 — Medium Impact

Issue B7: Replace 3×2 variant matrix with 4 product-tier CI matrix

Field	Value
Impact	Meaningful coverage per variant + text flag bug fix
Effort	3–4 hours

Bug found: The _multi_transition in rules.bzl:201 sets text_full but never sets //donner/svg/renderer:text. Since text defaults to false, all 6 transitioned variants run without text. The *_text and *_text_full suffixes produce identical binaries. The "6 variants" are really 3 (one per backend, all without text).

Also: The transitioned variants don't propagate shard_count, so each runs as a single enormous test process instead of 16 shards.

Proposed fix: Replace the 3×2 combinatorial matrix with 4 named product tiers that map to actual Donner build configurations:

Tier	Backend	Filters	Text	Bazel config	CI trigger
donner-tiny	tiny-skia	no	no	--config=tiny	All PRs
donner (default)	tiny-skia	yes	simple (stb)	--config=text	All PRs
donner-max	tiny-skia	yes	full (FT+HB+WOFF2)	--config=text-full	All PRs
legacy-skia-ref	skia	yes	full	removed full-Skia config + --config=text-full	main only

Changes needed:

Fix rules.bzl transition to also set //donner/svg/renderer:text=true.
Add --config=tiny to .bazelrc: common:tiny --//donner/svg/renderer:filters=false.
Replace the variants list in the resvg test BUILD with the 4-tier matrix.
Add shard_count propagation to donner_multi_transitioned_test.
Gate skia-ref variant to main-only (e.g. via a ci_full_matrix flag or separate target).

Key files: build_defs/rules.bzl, donner/svg/renderer/tests/BUILD.bazel, .bazelrc, .github/workflows/main.yml

Issue B8: Move coverage to nightly/merge-only schedule

Field	Value
Impact	Eliminates ~30 min job from every PR
Effort	10 min

Coverage runs on every push to every branch. Uses --nocache_test_results, full LLVM instrumentation (2× overhead), and requires JDK 11. Change trigger to push: branches: [main] + daily cron.

Issue B9: Enable ThinLTO + gc-sections for release builds

Field	Value
Impact	−10–20% binary size
Effort	15 min

No LTO configured anywhere. No -ffunction-sections -fdata-sections + -Wl,--gc-sections for Donner's own code (only for Skia's internal build).

Add to .bazelrc:

build:binary-size --copt=-flto=thin --linkopt=-flto=thin
build:binary-size --copt=-ffunction-sections --copt=-fdata-sections
build:binary-size --linkopt=-Wl,--gc-sections

Issue B10: Create forward-declaration headers for heavy types

Field	Value
Impact	−10–20% incremental compile time
Effort	3–5 days

Donner has zero *_fwd.h files. Every header that references Entity, Registry, or component types pulls the full definition chain. ~20 of 48 ECS includers only pass handles by value/reference and could use a lightweight forward declaration.

Issue B11: Split monolithic svg_core into focused libraries

Field	Value
Impact	~30% faster incremental builds
Effort	1–2 weeks

svg_core has 60 .cc files and 60 headers — shapes, filters, text, gradients, resources all in one target. Every consumer recompiles and links everything.

Proposed split:

Target	Contents (~files)	Rationale
svg_core_elements	Foundation types (~10)	Everything needs these
svg_core_shapes	Shape elements (~16)	Self-contained
svg_core_filters	FE* elements (~27)	Behind feature flag
svg_core_resources	Paint/resource elements (~15)	Paint servers

Issue B12: Unify coverage workflow cache key

Field	Value
Impact	Avoid redundant coverage rebuilds
Effort	5 min

Coverage uses bazelcov- prefix and hashes non-existent WORKSPACE* files. Never hits the main build cache.

Issue B13: Optimize Bazel cache key strategy

Field	Value
Impact	Higher cache hit rate
Effort	15 min

Current key includes .bazelrc — any comment change invalidates the entire cache. Split: use MODULE.bazel + .bazelversion for primary key, .bazelrc only in restore-key prefix.

Issue B14: Split monolithic renderer files

Field	Value
Impact	Better incremental compile times
Effort	3–5 days

The four largest .cc files are all renderers: the archived full-Skia renderer source (4,245 lines), RendererTinySkia.cc (2,254), RendererDriver.cc (2,072), RendererGeode.cc (1,896).

Split by concern: FullSkia_Gradient.cc, FullSkia_Filter.cc, FullSkia_Text.cc, FullSkia_Shape.cc. Changes to gradient code no longer recompile text code.

Build Improvement Impact Summary

Priority	Issues	CI Time Savings	Binary Size
P0	B1	−60–120 min/run	—
P1	B2–B6	−30–50% build+test	—
P2	B7–B14	−30 min/PR + incremental	−10–20%

Quick wins (< 1 hour total, immediate ROI):

B3: Increase --local_test_jobs (5 min)
B6: Add test sharding (30 min)
B9: Enable LTO + gc-sections (15 min)
B12: Fix coverage cache key (5 min)

Appendix: Spec Quick Reference

Area	Spec	Key Sections
Filter functions	Filter Effects 1	§8 (functions), §15 (feMorphology)
Transform origin	CSS Transforms 2	§6
Image sizing	SVG2 + CSS Images 3	SVG2 §5.8, CSS §5.2
Text baselines	CSS Inline 3	§4.2, §4.4
Paint order	SVG2	§13.11
Context paint	SVG2	§13.3
Switch	SVG2	§5.9
Masking	CSS Masking 1	§6.1, §7.2
BiDi	CSS Writing Modes 4 + UAX #9	§2
Text on path	SVG2	§11.7

Appendix: Bot Consultation Notes

This doc was synthesized from analysis by:

SpecBot — SVG2 spec section references, browser behavior cross-checks, priority ordering by test unblock count.
PerfBot — Measured CI run times, identified macOS queue bottleneck, EnTT template blast radius, compilation/binary size analysis.
BazelBot — Build graph analysis, svg_core split proposal, test sharding audit, cache key strategy, LTO/gc-sections recommendations.
Explore agents — Raw data extraction from resvg test suite, renderer gaps, CI workflows.

Each bot independently identified the same top priorities (CSS filter functions, transform-origin, EnTT modular headers, test sharding), giving high confidence in the recommendations.