tests: cover the QML protocol bindings and QML application use cases

[jcelerier]

Rebased on master and split into one commit per change. Depends on
ossia/libossia#939, now merged: the submodule pointer here is libossia master.

Fixes, and the test that covers each

commit	fix	covered by
build: asio buffer debugging	libossia sets BOOST_ASIO_ENABLE_BUFFER_DEBUGGING in Debug while the vendored liblsl does not, and the flag changes asio's internal layout — an ODR violation in any binary linking both	build-level; no runtime oracle
protocols: a mapper's script cannot outlive its device tree	queued slots and polling timers ran after stop() released the QQmlEngine; a script's resolved addresses were cached with no invalidation, so a device's destruction left dangling parameter_base* its onClose then wrote through	test_unit_mapper_protocols — "read and write callbacks survive repeated device removal", "an unqualified mapper address resolves in the script's own device"
js: a device identifier outlives the enumeration that made it	identifiers were deleted on every re-enumeration although the contract is to read them once and reuse their settings	test_integration_device_enumeration_lifecycle
js: an enumerator is a subscription, and a name is taken or it is not	the enumerator was parentless and collectible while still discovering; createDevice shadowed an existing name instead of refusing	same
js: one Qt Quick runtime per renderer, one publisher	the runtime was keyed on (thread, QRhi), so a second render list on the same RHI got the first one's runtime and tripped an assert; and several runtimes of one process each published its outputs	test_integration_texture_source_lifecycle
js: a destroyed texture source takes its preview node	unregistered through the item, which Qt has already destroyed by then, so the node and its render list stayed in the graph	same
avnd: a worker result is applied at the start of the node's own tick	results were applied before init_node cleared the outlets, with the previous tick's frame index, so objects had to re-emit from operator()	test_integration_avnd_worker_lifecycle
avnd: a CPU-only buffer producer is one object	it was classified as a GPU node and built once per RenderList, so two output windows meant two of whatever it owns	same

Tests

• The manual QML protocol corpus — 45 scenarios vendored under tests/testdata/protocols-qml, byte-identical to the originals apart from four import QtQml lines needed to parse, run by real Mapper devices over real sockets against independent peers. Framing, encoding and OSC oracles are written out in the tests, so score's encoder and decoder cannot agree on the same wrong bytes.
• Four use-case suites — texture source lifecycle, device enumeration lifecycle, Protocols.* from the --ui console engine, and the avendish worker/buffer contract.

Changes from the original single commit

• src/plugins/score-plugin-avnd/AvndProcesses/Test123.hpp removed: a scratch Distortion object, referenced by nothing.
• The submodule points at libossia master rather than the MIDI out apparently broken on mac #939 branch.
• Everything else is byte-identical to the original commit's tree.
• The four use-case targets were given the TIMEOUT 180 the three protocol targets already had, so a blocking test fails the run instead of holding it for the default twenty-five minutes.

Verification

Full build clean. 174 tests across the changed areas pass; the only failure is test_integration_js_remove_cable, which is red on master as well.

One case does not pass on a machine with a webcam, and it is not this branch's doing: test_integration_device_enumeration_lifecycle, "a device name already taken is refused rather than shadowed", stands two live Camera devices up and hangs on stop, in Video::CameraInput::close_file() joining its decode thread —

#7  std::thread::join
#8  Video::CameraInput::close_file   (CameraInput.cpp:240)
#9  Video::CameraInput::stop         (CameraInput.cpp:182)
#10 Gfx::video_texture_input_protocol::stop_execution
#11 Execution::DocumentPlugin::clear

close_file() does call m_interrupt.abort() before joining, so the buffer thread is blocked somewhere the interrupt callback does not reach — plausibly the v4l2 open of a second stream on a busy device. None of that code is touched by this branch, and the sibling case that stands up a single camera during playback passes. Reproducible 3/3 here. Worth fixing before this merges, since the new suite is the first thing to reach it.

🤖 Generated with Claude Code

https://claude.ai/code/session_019QWYTEhFJonuUNTEmBfv7Q