[jcelerier]

Clicking a combo box on the canvas did nothing — the only way to change one was
to drag it like a slider. This makes a click open the drop-down, keeping the
drag.

Revisits the QGraphicsCombo: open a popup menu on click commit that was left
out of #2262.

Why not the original approach

That commit called QMenu::exec() from mousePressEvent. That runs a nested
event loop while the item holds the implicit mouse grab — precisely the
situation the QEvent::UngrabMouse handling in the sliders, knobs and choosers
exists to survive — and self dangles if the process is deleted while the menu
is up. There is no other QMenu::exec() in a graphics item anywhere in score.

It also replaced dragging outright, which silently neutered
TEST_CASE("degenerate combo boxes survive being dragged") — the regression
test for a divide-by-zero on single-entry combos.

What this does instead

A press still begins a drag; the value only follows the pointer once it passes
QApplication::startDragDistance(). Below that it is a click, and a click opens
the same drop-down the right button already opened. Dragging is unchanged, its
regression test keeps its meaning, and a pixel of wobble while clicking no
longer nudges the value.

openEditor() already builds its editor from the event loop behind a
QPointer, with the index re-checked at commit time — no nested loop. Making it
reachable from both buttons needed two things of it:

• one drop-down at a time. The build is deferred, so a second click arriving
  before the first editor appeared would stack another on top; closing the pair
  frees the proxy twice.
• the pick is resolved by name, not by index. A runtime-populated combo box
  (the folder/dynamic comboboxes from Runtime-populated comboboxes and avnd UI file paths (reviewed #2110 + #2114) #2262) can be repopulated while the
  drop-down is open. The same index then means a different entry — or none, in
  which case nothing is selected rather than the wrong thing.

An empty combo box opens nothing.

Tests

Six new cases in GraphicsComboTest.cpp, driven through real scene events:
click opens / drag does not; three clicks plus a right click leave exactly one
editor; items reordered under an open drop-down still select what was picked;
an entry that has since disappeared selects nothing, leaving an index past the
end of the shorter list which is then painted to show readers tolerate it;
the combo box is deleted with the editor open and the editor activated anyway;
a degenerate combo box is clicked; and the scene taking the grab away mid-drag
is not mistaken for a click.

Verification

test_unit_graphics_combo: 67 assertions in 17 cases, all passing. Full
suite 145/150 — the five failures (analysis_gist, beat_tracker,
media_trim, unused_files, explorer_editor_look) reproduce on an unmodified
origin/master build.

Also run under valgrind: 17/17 pass, and no error context touches the combobox
code. The Qt-startup noise it does report (Invalid read of size 16 inside
QApplicationPrivate::init()) appears identically when running an untouched
pre-existing test, so it is environmental.

🤖 Generated with Claude Code

https://claude.ai/code/session_018oUFSPD1bswcWtvrsq25t3

[coveralls]

Coverage is 28.001% — feature/combobox-popup into master. No base build found for master.