WIP: UI smoke tests for axis, touchy, gmoccapy, qtdragon

[grandixximo]

Draft, opening for CI feedback. Refs #3756.

Summary

Phase 1 of the GUI test work tracked in #3756. Each test launches a GUI under xvfb-run against an existing configs/sim/<gui>/*.ini, drives Estop reset / machine on / home all via NML, asserts the interpreter reaches IDLE, then shuts down cleanly. Verifies the GUI starts and accepts basic commands without crashing.

Coverage

• axis
• touchy
• gmoccapy
• qtdragon (qtdragon_xyz/qtdragon_metric.ini)

Mechanics

• tests/ui-smoke/_lib/launch.sh: xvfb-run wrapper, setsid so the linuxcnc process group can be signalled cleanly, falls back to axis-remote --quit then SIGTERM with grace then SIGKILL. Skips with exit 77 if xvfb-run is unavailable (matches tests/tooledit and tests/pyvcp).
• tests/ui-smoke/_lib/drive.py: NML driver. Tolerant of sim configs that come up already in STATE_ON via auto-estop-release HAL wiring. Falls back to per-joint serial homing if no HOME_SEQUENCE is configured.
• tests/ui-smoke/_lib/checkresult.sh: pass when UI_SMOKE_OK printed and no crash markers in captured logs.
• Reuses existing sim configs, no test-only INI files.

Cleanup discipline

• .gitignore covers all runtime artifacts (linuxcnc.{out,err,pid}, ui-smoke.{out,err}, result, stderr)
• 4 consecutive runs locally: 4/4 pass, 0 shmem errors, working tree clean (no untracked files added beyond the committed test scripts). Aligns with the clean-tree gate Bertho asked for and that @hdiethelm is wiring up in CI improvemens: General improvements #3984.

Deps

xvfb is already declared in debian/control with the <!nocheck> profile so apt-get build-dep installs it on the existing CI without a workflow change. Coordinated with @hdiethelm in #3984: this PR adds no system deps; if his lands first, no rebase needed here.

Out of scope (deferred)

• Phase 2: load a small G-code file via linuxcnc.command.program_open + auto(RUN), verify final position via linuxcnc.stat.position. Per-GUI cross-checks via xdotool or AT-SPI where useful.
• Phase 3: screenshot or short video on failure, uploaded as CI artifact.

Test plan

• Local: 4/4 pass under scripts/runtests tests/ui-smoke, no shmem leaks
• CI: rip-and-test passes
• Reviewer feedback on scope: shipped smoke ("does it start, NML reachable, no startup crash"), per Bertho's framing in Add tests starting GUIs, likely falling back to xvfb for it #3756. Functional behaviour (load G-code, verify position) tracked as Phase 2 follow-up.

[hdiethelm]

Phase 3: screenshot or short video on failure, uploaded as CI artifact.

If you manage to create consistent screenshots and want to go to pedantic mode:

• Store reference known good screenshots (TBD where, I often use submodules for test data storage so the main repo is not overfilled and it is still tracked)
• Take screenshots at certain points where everything is static, like before / after homing / at the end
• Compare to the reference and highlight any differences, fail if there are differences -> Artifact
• The dev can download the artifacts, check them manually and if the change was on purpose replace the known good ones, so the CI passes again

Probably over complicated and I don't know how deterministic LinuxCNC is but this way, bugs like this #3979 can be easily avoided. Testing manually, these kind of bugs are just often overlooked.

[grandixximo]

Phase 3: screenshot or short video on failure, uploaded as CI artifact.

If you manage to create consistent screenshots and want to go to pedantic mode:

* Store reference known good screenshots (TBD where, I often use submodules for test data storage so the main repo is not overfilled and it is still tracked)

* Take screenshots at certain points where everything is static, like before / after homing / at the end

* Compare to the reference and highlight any differences, fail if there are differences -> Artifact

* The dev can download the artifacts, check them manually and if the change was on purpose replace the known good ones, so the CI passes again

Probably over complicated and I don't know how deterministic LinuxCNC is but this way, bugs like this #3979 can be easily avoided. Testing manually, these kind of bugs are just often overlooked.

The reference-screenshot diff approach is a good Phase 3 idea, will track it on #3756. For Phase 1 (this PR) I'm staying with NML state assertions only since they're deterministic; rendering will need the screen-stabilization tricks you mentioned.

[grandixximo]

Round 2 pushed; CI now passes 282/282 with all 4 ui-smoke tests running.

Changes in this round:

• NML connect-and-poll robustness (drive.py): linuxcncsvr's status buffer can be invalid (emcStatusBuffer invalid err=3) for the first ~30s after startup. Driver now retries the connect-and-poll cycle, recreating the stat object each iteration so a stale invalid buffer does not stick. CONNECT_TIMEOUT_S=60s.

• Software OpenGL (launch.sh): GitHub Actions runners have no GPU and qtdragon's GLcanon widget segfaulted under hardware GL. Set LIBGL_ALWAYS_SOFTWARE=1, GALLIUM_DRIVER=llvmpipe, QT_QUICK_BACKEND=software, QSG_RHI_BACKEND=software, QT_OPENGL=software.

• pyqt5-dev-tools Build-Depends: qtvcp compiles a QRC file via pyrcc5 at first run; without the package qtdragon segfaulted with "No such file or directory: 'rcc'". Added with <!nocheck> profile.

• Driver-trust checkresult (checkresult.sh): replaced the crash-marker grep with a simple UI_SMOKE_OK present + UI_SMOKE_FAIL absent check. The driver only prints UI_SMOKE_OK after a successful NML round-trip plus a re-poll after the GUI settle, so it is the authoritative signal. The previous regex was catching shutdown-side Qt teardown races that are out of scope for a startup smoke test.

• $HOME/linuxcnc/nc_files mkdir workaround: touchy's filechooser.py:29 does os.listdir($HOME/linuxcnc/nc_files) with no try/except and crashes on a clean $HOME. launch.sh pre-creates the directory as a workaround. Filing a separate issue for the underlying touchy bug.