[jcelerier]

Supersedes #2110 and #2114, which come from a fork and so could not take the
review fixes directly. Their commits are here unchanged, with authorship
intact; the four commits on top are the result of reviewing them, and the
avendish bump they depend on is first.

Depends on celtera/avendish#204 (merged), which supersedes celtera/avendish#171.

What the review found

The dynamic-combobox feature did not work at all. alternativesChanged was
declared with W_SIGNAL, which emits nothing for connections made from another
shared library — and every consumer is in one: the inspector and canvas
comboboxes live in score_lib_process's headers, the executor in
score_plugin_avnd. The item list would be replaced and nothing would redraw.
Every other signal on these ports already uses E_SIGNAL for exactly this
reason. The first new test catches it.

Folder-backed comboboxes listed nothing in ordinary documents. The sibling
folder port's value is a stored template (<PROJECT>:…, <LIBRARY>:…, or
document-relative) and was handed to QDir raw. Resolved now — via a walk up
the parent chain rather than IDocument::documentFromObject, which throws when
a process is not attached to a document yet.

#2114 resolved paths one way only. item.set writes straight back to the
inlet, so a custom widget echoing the absolute path it was just given would
store that path in the document and the project would stop being relocatable.
score's own file widgets have always paired locateFilePath with
relativizeFilePath; the other half is restored.

Extensions written as ".wav" matched nothing, becoming the glob "*..wav".
Both spellings are accepted now.

One commit from #2110 is not included. QGraphicsCombo: open a popup menu on click replaces drag-selection with QMenu::exec() — a nested event loop inside
mousePressEvent on a graphics item, which has no precedent in this codebase,
and which the QEvent::UngrabMouse handling in the sliders/knobs/choosers
exists precisely to avoid. It also silently neuters
TEST_CASE("degenerate combo boxes survive being dragged"), the regression test
for a divide-by-zero on single-entry combos. It is unrelated to the feature and
the remaining seven commits rebase cleanly without it, so it is left out; worth
its own PR if the popup behaviour is wanted.

On the DataStream change

ComboBox's DataStream layout gains two fields (folderPortName,
fileExtensions). I originally flagged this as a backwards-compatibility
hazard for .scorebin; that was wrong. The binary format is only used to
serialize in and out of a live score instance, so reader and writer are always
the same build and there is no older stream to desynchronise. No version guard
is needed. The JSON path is guarded with tryGet regardless.

Tests

tests/unit/ComboBoxAlternativesTest.cpp (7 cases, 41 assertions): the selected
value survives the list being replaced under it, falls back to init when it
disappears, and is left alone when it names a file an upstream process has not
written yet; folder listings are filtered, sorted, and degrade to a placeholder;
JSON round-trips preserve the new metadata and documents without it still load.

celtera/avendish#204 added the matching compile-time tests on the avendish side.

Verification

Built clean and run against master: 145/150. The five failures
(analysis_gist, beat_tracker, media_trim, unused_files,
explorer_editor_look) reproduce identically on an unmodified origin/master
build and are pre-existing. test_unit_graphics_combo still passes, which it
would not with the popup commit included.

🤖 Generated with Claude Code

https://claude.ai/code/session_018oUFSPD1bswcWtvrsq25t3

[coveralls]

Coverage is 27.983% — review/combobox-and-filepaths into master. No base build found for master.