fix: build bundled fcitx5 plugin against the bundled Qt

[xiaohan484]

Linked issue (required)

Fixes #5110

Summary / motivation (required)

CI builds the fcitx5 plugin against one Qt but ships it inside PyQt6's Qt — a
version mismatch that crashes Anki on startup for every fcitx5 user.

Root cause: the plugin subclasses QPlatformInputContext, a Qt private API
with no ABI guarantee across minor versions. Built against Qt 6.2.4 headers but
run inside Qt 6.11.0, it dereferences a misplaced member and SIGSEGVs on first
focus.

Solution: build the plugin against the same Qt as the bundled pyqt6-qt6
and install it where the packaging step (qt/tools/build_installer.py) looks.
The per-flag reasoning is in the commit message.

Steps to reproduce (required, use N/A if not applicable)

1. On a Linux desktop configured to use fcitx5 as the input method, install the
   official Anki 26.05 release.
2. Launch Anki normally.
3. It dies with SIGSEGV on startup, as soon as the main window takes focus.

How to test (required)

Checklist (minimum)

• I ran ./ninja check or an equivalent relevant check locally.
• I added or updated tests when the change is non-trivial or behavior changed.

Details

release.yml is workflow_dispatch-only, so it can't run on the PR. I
reproduced its build step locally: rebuilt the plugin against Qt 6.11.0
(confirmed the Qt_6.11 ELF tag and a clean dlopen(RTLD_NOW)), dropped it into
an Anki 26.05 bundle, and typed Chinese and Japanese without the crash.

Risk / compatibility / migration (optional)

1. Fixes build-linux-x86 only. build-linux-arm-installer installs a prebuilt
   fcitx5-frontend-qt6 from apt, built against Ubuntu's Qt rather than the
   bundled 6.11.0 — the same class of mismatch. I have no arm hardware to confirm
   or fix it, so it's left for a follow-up.
2. Adds an aqtinstall step (~1.5 GB Qt download) to the Linux release build.
3. Not validated on CI. release.yml is workflow_dispatch-only, so an
   external contributor can't run it; only its build step is reproduced locally
   (see How to test). Please confirm it in a real release build.

UI evidence (required for visual changes; otherwise N/A)

N/A

Scope

• This PR is focused on one change (no unrelated edits).

[iamllama]

Haven't tested yet, but a small nitpick: doubt we can rely on uv.lock's ordering here, maybe try parsing uv export --package aqt --extra qt's output instead?

[xiaohan484]

Good catch. Checked, and uv.lock has four pyqt6-qt6 entries — 6.11.0, 6.10.2,
6.8.1, 6.9.1 (one per extra). The original grep was picking one by ordering alone.

Switched to your suggestion:

qt_version=$(uv export --frozen --no-hashes --package aqt --extra qt \
  | grep -oP '^pyqt6-qt6==\K\S+')

Pushed as a follow-up commit. Thanks!

[iamllama]

Having re-read the linked forum post, building against qt6.2 doesn't seem to be the cause here?

[xiaohan484]

The thread establishes that deleting the .so stops the crash — so the plugin is what's crashing. It never gets to why; it stops at the workaround. So it isn't evidence against a Qt-version cause, it just never reaches one. Here's what I found when I went looking.

Qt's BC guarantee doesn't cover this plugin

Qt's docs:

There are no source or binary compatibility guarantees for the QPA classes, meaning that the API is only guaranteed to work with the Qt version it was developed against.

and the header itself, qtbase/src/gui/kernel/qplatforminputcontext.h:

This file is part of the QPA API ... Usage of this API may make your code source and binary incompatible with future versions of Qt.

Qt even installs it under a version-stamped private include dir — include/QtGui/6.11.0/QtGui/qpa/, reachable only via Qt6::GuiPrivate — which is why this PR has to pass -DQT_FIND_PRIVATE_MODULES=ON to compile at all.

fcitx5-qt is built directly on it — qt5/platforminputcontext/qfcitxplatforminputcontext.h (the qt6/ copy is a symlink to it):

 22: #include <qpa/qplatforminputcontext.h>
143: class QFcitxPlatformInputContext : public QPlatformInputContext {
216:     FcitxQtWatcher *watcher_;      // first data member

What actually breaks

qtbase d9bb8c0a17, "Automatically reflect new input context input direction when locale changes", first released in 6.7.0, adds both halves of this:

  // qplatforminputcontext.h
      friend class QInputMethod;
+
+     Qt::LayoutDirection m_inputDirection;

  // qplatforminputcontext.cpp
  QPlatformInputContext::QPlatformInputContext()
      : QObject(*(new QPlatformInputContextPrivate))
  {
+     // Delay initialization of cached input direction
+     // until super class has finished constructing.
+     QMetaObject::invokeMethod(this, [this]{
+         m_inputDirection = inputDirection();
+     }, Qt::QueuedConnection);
  }

That member takes sizeof(QPlatformInputContext) from 16 to 24.

The plugin is compiled against the 6.2 headers, where the base is still 16 bytes — so QFcitxPlatformInputContext's first data member, watcher_, lands at offset 16. In the 6.11 QtGui it is actually running against, offset 16 is m_inputDirection. The two overlap. And because the write is queued, it lands after the plugin's constructor has already set watcher_.

The result: QObject::connect(fcitx::FcitxQtWatcher, fcitx::FcitxQtInputContextProxy): invalid nullptr parameter, then SIGSEGV at 0x10. No keystroke needed — and that constructor is the frame in the forum backtrace.

There's a second, independent break, straight from the shipped 26.05 tarball:

$ cd anki-26.05-linux-x86_64/app_packages/PyQt6/Qt6

# what the bundled plugin (ELF tag Qt_6.2) needs:
$ nm -DC plugins/platforminputcontexts/libfcitx5platforminputcontextplugin.so | grep handleExtendedKeyEvent
     U QWindowSystemInterface::handleExtendedKeyEvent(..., QString const&, bool, unsigned short, bool)@Qt_6_PRIVATE_API

# what the Qt sitting next to it (6.11.0) provides:
$ nm -DC lib/libQt6Gui.so.6 | grep handleExtendedKeyEvent
0000000000264510 T QWindowSystemInterface::handleExtendedKeyEvent(..., QString const&, bool, unsigned short)@@Qt_6_PRIVATE_API

Both files come out of the same installer. The trailing bool was dropped after 6.2, so the symbol the shipped plugin imports does not exist in the Qt it is shipped with. It gets past dlopen only because the plugin isn't BIND_NOW.

Reproduction

On a machine running fcitx5 with no fcitx4 installed at all — so the fcitx4 angle raised in the thread isn't in play — QT_IM_MODULE=fcitx, which is what makes Qt load this plugin at all → SIGSEGV; unset, the plugin isn't loaded and there's no crash, the same effect as deleting it.

Rebuilding the same fcitx5-qt 5.0.10 source against 6.11, with nothing else changed → no crash, and CJK input works.

[read-the-docs-community]

Documentation build overview

📚 Anki | 🛠️ Build #33599320 | 📁 Comparing bf68cb0 against latest (cfe0077)

  🔍 Preview build  

4 files changed

± genindex.html
± autoapi/anki/collection/index.html
± autoapi/aqt/mediasrv/index.html
± autoapi/aqt/webview/index.html

[iamllama]

I confirmed that it builds the plugin properly, but was unable to reproduce the issue on a new ubuntu22 vm. fcitx5/mozc worked on there without it for me as well

@abdnh Thoughts?

[abdnh]

Thank you. Let's give it a test in the next beta.

Can you make a change to the CONTRIBUTORS file from the GitHub UI to fix the CI check?

[xiaohan484]

Done — updated it via the web UI, thanks!