[Yaraslaut]

Master CI has been red on most legs since #1004, and two nightly checks have failed three nights running (#1005, #1006). This branch fixes every cause found, including three library defects.

Library defects (include/morph, src)

• Use-after-free in ui::detail::Mounter. If a tab's page fails to mount, the failure is reported through onSelect, and that handler may destroy the Mounted. underOverlay's Restore then wrote into the freed Mounter. Valgrind reported Invalid write of size 8 … Restore::~Restore() (mount.hpp:244), reproduced locally on master. Only the page's mount now runs under the overlay, so the overlay is back in place before the failure is reported.
• Glaze brace-init builds one-element arrays. Json{std::string{text}} gave ["text"], so a string-typed field component reported its change as an array. Json{Json::object_t{}} gave [{}], so an MCP tools/call without arguments was refused. Both now use parentheses. Reverting either fix makes its new test fail.
• qmllint warnings in src/qt_quick.
  • MorphStack read Layout through children (typed QtObject), which Qt 6.8 rejects.
  • MorphPanel's divider set height on an item a layout manages.
  • Verified with Qt 6.8.1's own qmllint (the CI version): the original files reproduce CI's exact warnings, and the fixed files give none.

CI, tests and examples

• examples/client: saveGrab's parameter and kGrabSettle were unused without Qt Quick. That -Werror broke the asan, tsan and ubsan legs before any test ran, which is also why ASan never saw the use-after-free.
• test_check_install_export.sh: the dropped-detail-headers mutation is caught, but now on reactive/detail/graph.hpp, not completion_awaiter.hpp. Updated the needle. The installed tree itself was fine.
• test_sign_in (Windows): the temp path was spliced into JSON with backslashes. It now uses generic_string().
• test_qt_quick_table: the tests hardcoded a unit of 8 px, but the table pads by Theme.unitWidth, the font's average character width (6.08 px on CI). Also the const references clang-tidy asked for.
• cmake/tsan.supp: added race:std::logic_error::~logic_error, the logic_error form of the existing exception_ptr false-positive entries. A race-free standalone control reproduces the report on Linux/libstdc++. A planted race is still reported 10/10. Closes core: TSan reports a race between a Completion's exception being freed on the pool and read in onError on the main thread #985.
• test_push_net: the connect-and-reconnect test gets a 10 s budget. It timed out at 2 s under Valgrind on two of three nights.
• Coverage:
  • include/morph/ui fell to 91.37% against its 92% floor. New tests cover inline bytes and base64, field-component change typing, map-held rows, sizing validation and localized explorer titles. It now measures 92.16%.
  • include/morph/client gets a floor (4/4 arms).
  • Six moved allowlist hints in rational.hpp and document.hpp are updated.

Verification (local, Arch, GCC 16 / clang 23, Qt 6.12)

• gcc-debug: full ctest passed, 3991/3991, before the last two commits. Since then the affected [interpreter], [agent] and [explorer] tests pass there, and the coverage run below passes on the final tree. morph_tests [ui] under Valgrind: 0 errors.
• clang-coverage with CI's configure: 4131/4131 tests pass, and scripts/coverage.sh passes its gate.
• scripts/test_check_install_export.sh: all checks pass.
• examples/client/main.cpp compiles with -Werror, with Qt Quick both on and off.
• clang-format 22.1.8 is clean on every changed file. clang-tidy over the changed lines is clean (local 23, not CI's 22).
• Not verified locally: Windows, and the Valgrind and TSan nightly legs end to end. ci: nightly Valgrind memcheck failed #1005 and ci: nightly Linux / all optional features failed #1006 close themselves when their nightly jobs pass.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Hw9JacbWXaAPaoKGVy6ZHw

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!