You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Preflight fails on existing test issues — how should I submit fixes? #41
I've been working on a few fixes for the plugin and its renderer, but ran into some issues with the required preflight checks.
On my system (Arch Linux, Qt 6.12, Clang 23), the current upstream version fails several font and QML tests.
I've prepared separate fixes for both. When testing them together, all 4,269 backend tests, 889 SceneScript tests and 29 main CTest suites pass.
However, the full preflight still fails at LeakSanitizer, reporting around 2.7 KB of allocations at shutdown. The same allocation paths appeared in earlier runs, so this doesn't seem to be caused by my changes, although I haven't fully tracked down the cause.
I also ran into an issue with the pre-push hook and inherited Git environment variables when using worktrees.
I noticed that CONTRIBUTING.md requires a successful preflight before submitting PRs.
Would it be okay to open draft PRs with the existing failures documented, or would you prefer these issues to be resolved first?
I've kept the renderer and QML fixes separate, along with an unrelated Arch installer.
I've done some more digging and managed to narrow things down quite a bit.
The remaining LeakSanitizer findings seem to come mostly from external libraries:
Fontconfig: 895 bytes of global state at shutdown, stable even after 100 calls.
DBus/EGL: Around 1.6 KB originating from NVIDIA's EGL implementation. Exact ownership is still unclear.
NVIDIA GBM/EGL: A separate allocation growing by 136 bytes per lifecycle, reproducible without any wallpaper renderer code.
I've also reproduced these findings on unchanged upstream, so I haven't found evidence of a renderer ownership issue so far. I'd rather avoid adding cleanup calls to the plugin for something that appears to originate from external libraries.
On the positive side, I've prepared four separate contributions:
Renderer: Fixes the strict Release compilation warning and incorrect Fontconfig test assumptions.
QML tests: Fixes test isolation and integration issues with Qt 6.12.
Pre-push hook: Fixes inherited Git environment variables interfering with temporary test repositories. Added a regression test, now 18/18 passing instead of 17/18.
Arch installer: Adds a makepkg-based installer with backup handling and regression tests. removed due to irrelevancy (this was actually just for me to easier compile/install/switch the versions)
The renderer and QML fixes are independent. When applied together, all 4,269 backend tests, 889 SceneScript tests and 29 main CTest suites pass. The full preflight also gets through the strict builds, but eventually stops at LeakSanitizer (~2.7 KB in 14 allocations).
There's a bit of a catch-22, though: I can't push the fixes normally because the pre-push hook runs the full preflight, which already fails on the unchanged upstream baseline.
Everything is committed and ready locally. I'm planning to submit the changes as separate draft PRs, with the existing failures clearly documented rather than claiming a successful preflight.
Would you be okay with a one-time bypass of the local pre-push hook to publish those drafts?
I'll follow up with the PR links once they're available.
I had a look at the existing LeakSanitizer setup, and I think you're right. I'll focus on the preflight configuration rather than trying to fix allocations inside NVIDIA or Fontconfig.
I'll see if I can narrow the checks down without hiding actual bugs in the project's own code, and prepare a separate PR if I find a clean solution.
I'll also get the existing fixes up as draft PRs so you can review them.
I've managed to get the sanitizer checks working without completely disabling leak detection. GPU shutdown leaks are still reported, but no longer block the gate. ASAN/UBSAN errors and assertions remain fatal, and the other test suites still have strict leak detection.
With the renderer and QML fixes applied, all regular tests, strict builds, sanitizer checks and all 9 fuzz targets pass.
The preflight now exits with code 0, although Mull still can't run with LLVM 23, so no pass marker is generated.
Activity
CaptSilver commentedon Oct 20, 2026
Absolutely, we can hunt it down together.
VenRoot commentedon Oct 21, 2026
Thanks, I appreciate it!
I've done some more digging and managed to narrow things down quite a bit.
The remaining LeakSanitizer findings seem to come mostly from external libraries:
I've also reproduced these findings on unchanged upstream, so I haven't found evidence of a renderer ownership issue so far. I'd rather avoid adding cleanup calls to the plugin for something that appears to originate from external libraries.
On the positive side, I've prepared four separate contributions:
Arch installer: Adds a makepkg-based installer with backup handling and regression tests.removed due to irrelevancy (this was actually just for me to easier compile/install/switch the versions)The renderer and QML fixes are independent. When applied together, all 4,269 backend tests, 889 SceneScript tests and 29 main CTest suites pass. The full preflight also gets through the strict builds, but eventually stops at LeakSanitizer (~2.7 KB in 14 allocations).
There's a bit of a catch-22, though: I can't push the fixes normally because the pre-push hook runs the full preflight, which already fails on the unchanged upstream baseline.
Everything is committed and ready locally. I'm planning to submit the changes as separate draft PRs, with the existing failures clearly documented rather than claiming a successful preflight.
Would you be okay with a one-time bypass of the local pre-push hook to publish those drafts?
I'll follow up with the PR links once they're available.
CaptSilver commentedon Oct 23, 2026
That sounds good.. the fix might be in preflight itself. I am not trying to sanitize the whole stack.
VenRoot commentedon Oct 29, 2026
Ah, that makes sense!
I had a look at the existing LeakSanitizer setup, and I think you're right. I'll focus on the preflight configuration rather than trying to fix allocations inside NVIDIA or Fontconfig.
I'll see if I can narrow the checks down without hiding actual bugs in the project's own code, and prepare a separate PR if I find a clean solution.
I'll also get the existing fixes up as draft PRs so you can review them.
VenRoot commentedon Oct 29, 2026
I've managed to get the sanitizer checks working without completely disabling leak detection. GPU shutdown leaks are still reported, but no longer block the gate. ASAN/UBSAN errors and assertions remain fatal, and the other test suites still have strict leak detection.
With the renderer and QML fixes applied, all regular tests, strict builds, sanitizer checks and all 9 fuzz targets pass.
The preflight now exits with code 0, although Mull still can't run with LLVM 23, so no pass marker is generated.
I've opened four separate draft PRs:
I kept everything separate to make reviewing easier. The remaining limitations and test results are documented in each PR.
Thanks again for pointing me in the right direction!