Design doc for upgrading Donner's vendored resvg-test-suite from the 2022-02-12 snapshot we currently pin to the current upstream HEAD (late 2024), handling the 2023-05-05 restructure, the 2023-05-06 "Great Rename", the repo move from RazrFalcon to linebender, and ~2.5 years of accumulated test additions and golden updates — without losing the ~220 hand-triaged override entries that encode everything we know about feature gaps, bugs, and spec-vs-resvg disagreements.
Motivation
Our pinned commit is 2.5 years old and missing hundreds of new tests covering features we care about (filter primitives, text, marker, mask, paint servers). Upstream has also fixed golden bugs that we currently work around with custom goldens. Staying pinned means:
- New SVG features ship without regression coverage
- Renovate PRs accumulate in a "too hard to land" backlog (renovate/resvg-test-suite-digest has been floating for a while)
- Our golden overrides may be masking bugs upstream has since fixed
- Every new contributor asks "why don't our tests match upstream's naming?"
The upgrade is large but it is strictly better than a slow rebase: we get everything in one PR, one triage pass, one set of goldens to re-bless.
Current State
Vendoring
Test integration
- Entry point: donner/svg/renderer/tests/resvg_test_suite.cc
- Discovery: getTestsWithPrefix(const char* prefix, ...) at resvg_test_suite.cc:14 — iterates @resvg-test-suite//svg (non-recursive) and matches by filename prefix (a-fill, e-textPath, etc.)
- Golden pairing at resvg_test_suite.cc:48 — <svg stem>.png in @resvg-test-suite//png/, unless overridden by Params::overrideGoldenFilename
- Test suite instantiation: ~74 INSTANTIATE_TEST_SUITE_P blocks across ~28 prefix groups
- Canvas size: forced to 500×500 for all tests via params.setCanvasSize(500, 500) at resvg_test_suite.cc:34
Override inventory (current)
| Override | Count | Meaning |
| Params::Skip() | ~77 | Feature not implemented, or known bug we haven't rooted out |
| Params::RenderOnly() | ~51 | Renders fine, pixel compare not meaningful (includes the 49 UB tests from jwmcglynn/donner#496) |
| Params::WithThreshold(...) | ~89 | AA artifacts, minor rendering diffs |
| .onlyTextFull() | ~15 | Needs HarfBuzz backend (--config=text-full) |
| Params::WithGoldenOverride(...) | ~40 | resvg's golden is wrong per spec; stored in donner/svg/renderer/testdata/golden/resvg-*.png |
Total test entries: roughly 1087 lines in resvg_test_suite.cc, with ~220 of them carrying non-default params. Each one represents a judgment call that took real time and must survive the migration.
Custom goldens
- Location: donner/svg/renderer/testdata/golden/resvg-<test-name>.png
- Documented in 0009-resvg_test_suite_bugs.md with per-bug Symptom → Root Cause → Ruled Out → Resolution format
- ~40 files; named using the old a-* / e-* convention
Triage tooling
tools/mcp-servers/resvg-test-triage/ — MCP server with analyze_test_failure, batch_triage_tests, detect_svg_features, suggest_skip_comment, and generate_feature_report. All of its pattern matching currently assumes the old a-* / e-* naming convention and will need to be updated in lockstep.
Upstream State
Repo move
- New canonical URL: https://github.com/linebender/resvg-test-suite
- Default branch renamed master → main in 3c2309ffde (2024-10-29)
- Old URL still redirects, but the Bazel new_git_repository pin should point at the new URL to avoid Renovate confusion
Key commits to migrate across
| Commit | Date | What it does |
| 7dea0ef591 | 2023-05-05 | Restructure tests. Flat svg/ + png/ → hierarchical tests/<category>/<feature>/<NNN>.svg with paired .png in the same leaf directory. a-/e- prefix dropped — the directory tree now encodes that distinction. images/ → resources/. |
| dd23ec1319 | 2023-05-06 | The great rename. Numeric NNN.svg → descriptive kebab-case (e.g. 001.svg → accumulate-with-new.svg). Goldens renamed alongside. No mapping document — only git's rename detection tells you old → new. |
| 58104c8752 | 2023-05-06 | Rename more. Cleanup of the great rename. |
| 9b6f7d9f39 | 2023-05-06 | Remove dummy tests. Some of our entries will resolve to "deleted upstream". |
| 2023-05-15..2023-05-21 | | New filter/mask/text/clipPath/stroke-dasharray tests added |
| 2023-10-01 | | More marker/fill/svg tests; reference image updates |
| d20d47b514 | 2024-08-20 | Sync tests with resvg. Large sync pulling in upstream changes |
| 3c2309ffde | 2024-10-29 | master → main rename |
| d8e064337f | 2024-10-29 | HEAD as of this design doc. 1679 .svg files under tests/, 7 top-level categories (filters/, masking/, paint-servers/, painting/, shapes/, structure/, text/). |
New layout sample
Before (what our code assumes):
svg/
a-fill-001.svg
a-fill-010.svg
e-filter-013.svg
png/
a-fill-001.png
a-fill-010.png
e-filter-013.png
images/
...
After (current upstream):
tests/
painting/
fill/
#RGB-color.svg # was a-fill-001.svg
#RGB-color.png
rgb(int int int).svg # was a-fill-010.svg (the UB test)
rgb(int int int).png
...
filters/
enable-background/
new.svg # was filters/enable-background/009.svg
accumulate-with-new.svg
with-clip-path.svg
...
resources/ # was images/
fonts/
Note the descriptive filenames include characters (#, (, ), spaces) that can be awkward in C++ test names — we'll need to sanitize them (GoogleTest already does this for TestNameFromFilename; verify it still produces stable identifiers).
Goals
- Update to current upstream HEAD (pinning to the commit, not a branch)
- Map every current override from old name to new name, preserving intent
- Surface orphans: entries whose tests were deleted upstream
- Surface new tests: file list of .svg's not previously covered, for triage in a follow-up PR
- Keep the a-/e- prefix-based test grouping conceptually, translating it to directory walks — contributors already think in those buckets
- Verify custom goldens still apply — some may now be unnecessary because upstream fixed its goldens; others need their filenames updated
- Update the triage MCP server so future failures get the right feature labels
Non-goals
- Triage new tests in the landing PR. A few hundred new .svg files will have no entries. They'll use default params and most will fail. The landing PR gets a conservative Skip/RenderOnly sweep just to turn CI green; the real triage is tracked as a fast-follow milestone (see Work Breakdown below), not something that drifts into the backlog.
- Switch to linebender's result tracking (results.csv, stats.py, site/svg-support-table.html). Interesting, but scope creep.
- Adopt upstream check.py pre-commit hook. Ours are different.
- Fix underlying rendering bugs exposed by the upgrade. Each bug becomes its own issue, triaged alongside the new-test enabling work.
Migration Strategy
Phase 0: Build the rename mapping (one-time, offline)
The hardest part. No upstream migration document exists; the only source of truth is git's rename detection across the two restructuring commits.
Approach: clone linebender/resvg-test-suite locally, run a script that:
- Walks every .svg in the old pinned commit (682a9c8) listed in our resvg_test_suite.cc overrides
- For each, runs git log --follow --name-status --diff-filter=R
682a9c8..d8e064337f -- <old-path> to collect all rename events
- Produces tools/resvg_test_suite_upgrade/rename_map.json mapping old_filename → new_relative_path (or null for deletions)
Why --follow rather than diffing the tree commits directly: the restructure and great rename are two separate commits, plus 58104c8752 "Rename more." and some ad-hoc moves afterwards. --follow chains all of them.
Caveats:
- --follow only tracks one path at a time — the script has to loop.
- Git rename detection is similarity-based; files that were also edited in the same commit may show as delete+add instead of rename, requiring --find-renames=50% tuning.
- Some tests will have been deleted (9b6f7d9f39 Remove dummy tests.) and show as null. That's expected; the script should report them as "orphaned".
Validation: spot-check 20 entries manually against the GitHub commit history, especially ones from filters/enable-background/ where we saw the 001 → new-with-region style reshuffling (not order-preserving).
Phase 1: Update vendoring
- Update third_party/bazel/non_bcr_deps.bzl:113 to pin d8e064337f (or whatever is HEAD at PR time) and change the remote to https://github.com/linebender/resvg-test-suite.git
- Rewrite third_party/BUILD.resvg-test-suite to match the new layout:
filegroup(
name = "tests",
srcs = glob(["tests/**/*.svg", "tests/**/*.png"]),
visibility = ["//visibility:public"],
)
filegroup(
name = "resources",
srcs = glob(["resources/*"], allow_empty = True),
visibility = ["//visibility:public"],
)
filegroup(
name = "fonts",
srcs = glob(["fonts/*.{ttf,otf}"], allow_empty = True),
visibility = ["//visibility:public"],
)
- Update BUILD.bazel test targets at donner/svg/renderer/tests/BUILD.bazel:295-301 to reference the new filegroups (tests replaces both svg and png).
- Delete the :svg, :png, :images filegroup names — with the flat layout gone, they no longer make sense. Prefer a single :tests filegroup.
Phase 2: Rewrite discovery
getTestsWithPrefix(prefix) at resvg_test_suite.cc:14 currently does a non-recursive scan and prefix match. Two options for the new layout:
Option A — Walk by directory (preferred)
Replace the prefix API with a directory-scoped API:
std::vector<ImageComparisonTestcase> getTestsInCategory(
std::string_view category,
std::map<std::string, Params> overrides = {},
Params defaultParams = {});
Callers become:
INSTANTIATE_TEST_SUITE_P(Fill, ImageComparisonTestFixture,
ValuesIn(getTestsInCategory("painting/fill", { ... })),
TestNameFromFilename);
Pros: mirrors the new directory tree 1:1; no prefix ambiguity; cheap recursive_directory_iterator on exactly one subtree.
Cons: a-fill and a-fill-opacity used to be separate prefixes; now they're painting/fill/ and painting/fill-opacity/ — same outcome, so this is actually cleaner.
Option B — Keep prefix API but compute synthetic prefix from the new path (painting/fill/foo.svg → synthetic a-fill-foo). Lets the rename mapping be used to keep override keys stable.
Decision: Option A. The synthetic-prefix bridge is clever but leaks the old layout into the new code forever. Better to bite the bullet once.
Override-map keys become the new filename (e.g. "rgb(int int int).svg" — yes, with parens). The rename map JSON from Phase 0 feeds a codemod that rewrites every {"a-fill-010.svg", ...} entry to its new form.
Phase 3: Codemod the test entries
Write tools/resvg_test_suite_upgrade/rewrite_test_entries.py that:
- Parses resvg_test_suite.cc with a line-oriented regex that matches the {"a-foo.svg", Params...}, // comment entries — the file is mechanically structured, so tree-sitter is overkill
- Looks up each old filename in the rename map
- Emits new entries with the new filename and preserved comment
- Outputs a report:
- Migrated: N entries
- Orphaned (deleted upstream): list, each needs a decision
- Ambiguous (rename map has multiple candidates): list
- Groups the new entries by category directory and regenerates the INSTANTIATE_TEST_SUITE_P blocks in the order of the new tree
Hand review of orphaned entries: each one used to encode a reason (Skip/threshold/bug comment). For each, either:
- Verify the test really is gone upstream and drop the entry, OR
- Find the new test that replaced it and re-target the override
Phase 4: Handle new tests
Once the codemod lands, bazel test will enumerate all ~1679 tests. The ones with no override will use default params, and many will fail. Strategy:
- Run the suite once, collect failures into new_test_failures.txt
- Use the resvg-test-triage MCP server (updated in Phase 6) to bulk-categorize them by feature
- Add override entries in a single bulk commit per category, using conservative defaults:
- Obvious "not impl" → Params::Skip() with // Not impl: <feature>
- Passes with small diff → Params::WithThreshold(0.1f)
- Large diff but renders → leave failing, add to triage backlog (do not auto-skip; skip/render-only needs human approval per project policy)
- Triage is a follow-up effort, not part of the landing PR. The landing PR should end with a clean bazel test //... but may have many more entries marked Skip than before.
Phase 5: Custom goldens
The 40 files in donner/svg/renderer/testdata/golden/resvg-<old-name>.png need:
- Renaming to match the new test names (via the rename map)
- Re-validation: for each, diff against the new upstream golden. If upstream has since fixed the bug (e.g. marker direction, filter blur halo), delete the custom golden and trust upstream.
- Update 0009-resvg_test_suite_bugs.md entries to reference the new names and note any that were retired.
- Update Params::WithGoldenOverride(...) paths in resvg_test_suite.cc to the new filenames.
Phase 6: Update triage tooling
tools/mcp-servers/resvg-test-triage/ assumes a-* / e-* naming. Update:
- Pattern matching in detect_svg_features — switch from filename prefix to directory-based category detection
- Feature report generation in generate_feature_report — group by new category tree
- Skip comment templates in suggest_skip_comment — no longer needs to emit a-* style paths
Phase 7: Land
Final PR contents:
- third_party/bazel/non_bcr_deps.bzl — new commit + URL
- third_party/BUILD.resvg-test-suite — new layout
- donner/svg/renderer/tests/BUILD.bazel — updated data deps
- donner/svg/renderer/tests/resvg_test_suite.cc — fully rewritten with new entries, grouped by new categories
- donner/svg/renderer/testdata/golden/ — renamed custom goldens (or deletions where upstream fixed things)
- docs/design_docs/0009-resvg_test_suite_bugs.md — updated
- tools/mcp-servers/resvg-test-triage/ — updated
- tools/resvg_test_suite_upgrade/ — keep the rename map JSON + scripts as historical artifacts (useful if we ever need to re-run)
- MEMORY.md note updates: test filter syntax, resvg test target location
Risks
R1: Rename map is incomplete or wrong
Impact: Entries silently drop their override; tests that used to Skip now fail.
Mitigation: the codemod's "migrated count" must equal "old override count". Any discrepancy fails the script. Manual spot-check of at least 20 entries. Keep the rename map checked in so it's reviewable.
R2: Upstream fixed something we have a golden override for, and our override now masks a Donner regression
Impact: silent loss of coverage.
Mitigation: in Phase 5, diff each custom golden against both Donner's current output and upstream's new golden. If upstream matches Donner, delete the custom golden. If upstream matches our custom golden but differs from current Donner output, that's a Donner regression we need to investigate before landing.
R3: New tests expose a large crop of bugs that can't all be triaged in one PR
Impact: PR either lingers unmerged, or lands with hundreds of fresh skips.
Mitigation: explicit non-goal — triage is follow-up work. Landing criterion is "bazel test //... green", not "every test passing". The post-landing state has more skips; that's the honest representation.
R4: File names with special characters (#, (), spaces) break tooling
Impact: GoogleTest name sanitization produces unstable IDs, or Bazel globs miss files, or shell commands in scripts mis-parse.
Mitigation: verify TestNameFromFilename produces stable output for rgb(int int int).svg. Bazel globs should be fine (they pass filenames verbatim). Scripts must quote paths properly — write in Python, not bash.
R5: Renovate PR conflicts mid-migration
Impact: Renovate opens a PR pointing at an even newer HEAD while our branch is in flight.
Mitigation: close the renovate PR explicitly before starting; rebase to the latest HEAD just before merging.
R6: Lost triage context
Impact: Comments on entries (// Bug: Kerning on textPath) are the only record of why something is Skip'd. If a comment doesn't survive the codemod, we lose institutional memory.
Mitigation: codemod preserves the trailing // ... comment verbatim. Validated by a round-trip test (parse → rewrite with identity mapping → diff should be empty).
R7: The upgrade exposes that several Params::Skip() entries were masking actual bugs
Impact: can't land the PR without also fixing bugs.
Mitigation: per project policy, we never skip tests to make CI green. Any newly-exposed bug becomes a blocker issue and gets its own fix PR before the upgrade lands. This is the correct outcome, even if it delays things.
Open Questions
- Do we want to pin to a tag or to HEAD? Upstream has no tags since 2018. Pinning to a specific commit is the only option, and Renovate will keep opening PRs. Fine.
- Should we keep the svg/ + png/ flat layout in our overlay BUILD file by symlinking, to minimize C++ changes? Tempting but wrong — it hides the upgrade and makes future debugging harder. Do the real migration.
- Are the resources/ files referenced by xlink:href inside SVGs? If an SVG does <image href="../../../resources/foo.png"/>, the runfiles path has to be right. Need to check whether any test SVG uses relative asset refs and verify the Bazel filegroup layout supports them. Likely need resources as a sibling of tests/ in runfiles.
- What to do about images/ (old) assets that referenced tests may still point to? If our pinned commit's SVGs reference images/foo.png and the new tree has those as resources/foo.png, pre-migration tests would break — but we're not doing an intermediate step, we're going straight to new HEAD, so this only matters if any individual SVG internally hardcodes the path. Audit needed.
- Goldens for the new Skia backend: do we need to regenerate -skia suffixed goldens too, or are they under the same naming scheme? Check the legacy full-Skia golden handling.
- Should the design include upgrading resvg-test-triage to a language model-assisted triage flow (e.g., feed failures to Claude with the SVG and diff image)? Out of scope for this doc but worth noting as a natural follow-up.
Work Breakdown
Two milestones. Milestone 1 is the landing PR — large but scoped to the mechanical migration. Milestone 2 is the fast-follow that actually enables the new tests we just pulled in. Both are required for the upgrade to be "done"; skipping M2 means we paid the migration cost without getting the coverage benefit.
Milestone 1 — Mechanical migration (landing PR)
Rough sequencing; each bullet is a self-contained unit of work within one PR.
- Prep: clone upstream locally, verify git log --follow works across the restructure commits on a single file as a smoke test
- Phase 0: write & run rename-map script; check in tools/resvg_test_suite_upgrade/rename_map.json + script
- Phase 1: vendoring update + BUILD file rewrite (file compiles, tests still enumerate old entries — they'll fail, that's fine locally)
- Phase 3: codemod the C++ test file (driven by rename map); validate orphan list manually
- Phase 2: rewrite getTestsWithPrefix → getTestsInCategory
- Phase 5: custom golden re-validation + rename
- First green test run: bazel test //... — many new failures expected
- Conservative sweep for new tests: enough Skip/threshold entries to get CI green, nothing more. No feature-level triage yet.
- Phase 6: triage MCP server updates
- Phase 7: land the PR (expect heavy review; it's a big diff)
Exit criterion: bazel test //... green on main. Probably hundreds more Skip entries than before — this is acknowledged and the fix is Milestone 2.
Milestone 2 — Enable new tests (fast-follow)
Kicks off the day M1 lands. Goal: work through every newly-added test and either enable it (with a threshold if needed), flag it as a real bug, or confirm it's a feature we don't implement yet. Each category is its own PR — small, reviewable, and safe to ship independently.
- Generate the new-test inventory: diff the post-migration test set against the pre-migration set; group by category directory. Produces a tracking issue with a checklist per category.
- Run the triage MCP server across the new-tests inventory to auto-categorize failures (not impl / threshold / bug / font difference)
- Per-category enablement PRs, in roughly this priority order (pick whatever's highest-value first):
- painting/fill, painting/stroke-*, painting/fill-rule — core paint features, likely the fewest surprises
- shapes/* — basic shape tests, should mostly "just work"
- structure/* — <svg>, <g>, <use>, <symbol> edge cases
- paint-servers/* — linear/radial gradient, pattern
- masking/* — clip-path and mask
- filters/* — largest and most likely to expose bugs; probably multiple PRs per filter primitive
- text/* — most complex, run across both text simple and text-full backends per project convention
- For each enabled test that exposes a real bug: open a tracking issue, leave the test failing or skipped with a link to the issue, and keep moving. The goal of M2 is coverage, not bug fixing — bug fixing happens in dedicated PRs afterwards.
- Milestone complete when: every new test has an explicit status (passing / threshold / bug-linked skip / not-impl skip) and the project Skip count is back down to "known gaps only".
Milestone 3 — Ongoing (not scoped here)
- Incremental bug fixes against the tracking issues created in M2
- Renovate auto-bumps for future upstream updates (should be painless once we're on the new layout)
Extra Credit — Publishing Donner to the upstream support table
Once M2 is done and we're sitting on a real pass rate, we should publish Donner as a column on linebender.org/resvg-test-suite/svg-support-table.html, the same table that shows resvg, Chrome, Firefox, Safari, Batik, Inkscape, librsvg, svgnet, and QtSvg. That's the canonical place the SVG rendering community looks for "how good is engine X at SVG?" — we want Donner on it.
How the upstream table actually works (it's not what you'd expect)
The support table is human-triaged, not automated. The workflow is:
- A Qt5/6 GUI tool at tools/vdiff/ in linebender/resvg-test-suite shells out to each renderer binary per test SVG, collects the rendered PNG, and shows a side-by-side "reference vs all backends" view.
- A human clicks Passed / Failed / Crashed / Unknown per test per backend.
- Scores land as integers (0=unknown/UB, 1=passed, 2=failed, 3=crashed) in results.csv, one column per backend.
- site/gen-html-tables.py + stats.py regenerate the published HTML table and chart.svg from that CSV.
There is no automated submission API, no CI integration, no pixel-diff equivalence check. It's a shipping spreadsheet.
Ladybird's precedent: PR #55, still draft after 17 months
The best prior art for what we'd be doing is linebender/resvg-test-suite#55 — "Test Ladybird", by shlyakpavel (not a Ladybird maintainer). It's a draft PR against the vdiff tool that:
- Adds a Ladybird entry to the Backend enum in tools/vdiff/src/tests.h
- Implements renderViaLadybird in tools/vdiff/src/render.cpp (~30 LoC; shells out to headless-browser --screenshot=1 --width=500 --height=500 foo.svg and reads output.png)
- Adds a ladybirdPath setting + file picker in settings.{h,cpp} and the settings dialog .ui forms
- Appends a ladybird column to results.csv and runs the triage manually
Maintainer RazrFalcon responded: "That's the easy 25%. Not worth adding to
the main test suite yet. Let's wait for at least 50–60%." The PR stalled at 27.5% passing. No updates since November 2024. Lesson: don't bother opening the PR until we're ≥60%.
Ladybird also has a long-open tracking issue LadybirdBrowser/ladybird#2306 for the integration. The only Ladybird-side change that shipped to support the effort was LadybirdBrowser/ladybird#2323 — adding --width/--height flags to headless-browser to get 500×500 output matching the suite's canvas.
Proposed Donner integration
Two separate deliverables, both in the Extra Credit bucket:
E1: Donner-side CLI for rendering single SVGs at 500×500
Make sure donner_svg_tool (or a sibling binary) can render a single SVG to a PNG at 500×500 with the right flags:
donner-svg-tool render --width=500 --height=500 --output=/tmp/out.png foo.svg
This is mostly already present — verify the surface area, add --width / --height if missing, ensure exit code is nonzero on crash (not just on parse error), and stabilize the output path so vdiff can find it.
Also add a flag to select the backend (--backend=skia / --backend=tiny-skia) so we can submit one column per backend, or merge them into the best result. Probably cleaner to submit two columns — donner-skia and donner-tiny-skia — so the table reflects reality.
E2: Upstream PR against linebender/resvg-test-suite
Following Ladybird's pattern in linebender/resvg-test-suite#55:
- Add Donner (and DonnerTinySkia?) entries to the Backend enum in tools/vdiff/src/tests.h — BackendsCount increments, which cascades into the hand-maintained .ui forms. Budget the pain.
- Implement renderViaDonner in tools/vdiff/src/render.cpp:
QStringList arguments = {
"render",
"--width=500", "--height=500",
"--output=" + Paths::workDir() + "/output.png",
data.imgPath
};
Process::run(data.convPath, arguments, true);
- Add donnerPath field + settings dialog entries for the binary picker.
- Locally run vdiff across all ~1680 tests, triage Passed/Failed/Crashed for every SVG. This is hours of clicking; budget a quiet afternoon.
- Append donner column to results.csv with those triaged scores.
- Open the PR. Expect maintainer pushback if we're under 60% passing.
Gating: don't spend time on E2 until M2 has driven the automated pass rate above 60% of non-UB tests (matches the bar RazrFalcon set on Ladybird's PR). Our automated pixel-diff count will be an over-estimate of manual Passed (pixel-diff fails AA artifacts; manual triage accepts them), so the manual vdiff score will probably be higher than our pixel pass rate. But the automated number is the cheap proxy to track.
Non-goals for Extra Credit
- Don't rewrite vdiff to use Donner's pixel-diff harness. vdiff exists as shared community triage infrastructure; changing its model is out-of-scope for Donner.
- Don't try to automate the PR (e.g., a bot that re-runs + updates the CSV). Other backends don't do this, maintainers won't accept it.
- Don't block M1 or M2 on Extra Credit. It only makes sense after M2 gives us a stable ≥60% pass rate.
Risk: the bar keeps moving
If we hit 60% passing and open the PR, the maintainers might say "come
back when you're at 80%." Budget for multiple rounds. The value of landing in the table is community recognition, not a one-time thing.
Appendix: Rename Map JSON Format
{
"a-fill-001.svg": "tests/painting/fill/\#RGB-color.svg",
"a-fill-010.svg": "tests/painting/fill/rgb(int int int).svg",
"e-filter-034.svg": "tests/filters/filter/in=FillPaint.svg",
"a-some-removed-test.svg": null
}
The script that generates it must:
- Include every .svg referenced in resvg_test_suite.cc overrides
- Additionally include every .svg that existed in 682a9c8 (to catch tests that currently use default params — they might still be worth knowing about)
- Use git rename detection with --find-renames=50% and log any matches below 80% similarity for manual review
Related Docs
1.15.0