[abrichr]

Why

The release gate (scripts/check_release_ci.py) has required a live-qualification.yml run on a self-hosted, qualified Linux desktop since #132 (Sep 3). That runner was never registered. CAPTURE_SELF_HOSTED_QUALIFIED_LINUX_RUNNERS is unset, and the repo and org both have 0 runners. So no release could pass. The 1.3.0 release run 39075912841 failed closed on exactly this. Richard approved making the GitHub-hosted Ubuntu Xvfb lane the release gate.

What changes

• New job in production-qualification.yml: Hosted Linux live qualification (ubuntu-latest, Xvfb). It installs the exact candidate wheel with [linux] into an isolated venv. Then, inside one dbus-run-session, it starts:

  • Xvfb 1280x1024x24 with RECORD and XTEST (both checked with xdpyinfo)
  • at-spi-bus-launcher and openbox
  • a focused GTK Entry window (relaunched for each trial)

  It runs three counted trials of tests/test_structural_observation.py and tests/test_performance.py -m slow, which is 8 tests per trial. That covers the recorder lifecycle, pynput→XTEST→X RECORD input capture, memory/CPU bounds, and AT-SPI structural observation. Each trial must pass check_junit_no_skips.py --expected-tests 8, and then aggregate_qualification_trials.py --expected-trials 3 must pass. The evidence is uploaded as an artifact.

• check_release_ci.py: the new job joins EXPECTED_QUALIFICATION_JOBS. The live-qualification.yml requirement and its Linux-job filtering are removed. Every release requirement now runs on GitHub-hosted runners.

• live-qualification.yml is unchanged in behaviour. It still runs weekly on its schedule and on dispatch, but it is documented as outside the release gate.

• Tests, release.yml step names, docs/DESIGN.md, docs/WINDOW_CAPTURE.md, and the v1.3.0 CHANGELOG are updated to match.

What the gate no longer proves (honest scope)

The hosted lane is weaker than the self-hosted lane was meant to be. A release no longer proves:

• more than one monitor. MSS on Xvfb reports a single monitor, so the topology check uses --minimum-monitors 1.
• real input hardware. Input is injected through XTEST, not a physical keyboard or mouse.
• a real compositor, real window moves, or a real desktop session.

The weekly non-blocking live-qualification.yml lanes remain the place for those, once a qualified desktop exists.

Verification

• I extracted the exact run: block of the trial step from this YAML and ran it on a Debian 13 box with the 571d530 wheel. All 3 trials passed with 8 passed, 0 skipped each. Each trial passes check_junit_no_skips.py --expected-tests 8, and the aggregate reports passed: true. An earlier 1920x1080 screen exceeded the 500 MB memory bound in test_memory_bounded (511.6 MB), so the lane uses 1280x1024. The threshold is unchanged.
• uv run --extra dev pytest -q tests: 1036 passed, 22 skipped. ruff check openadapt_capture/ is clean. scripts/check_changelog.py is OK.

Signed-off-by: Richard Abrich richard.abrich@mldsai.com

[abrichr]

Self-review before merge:

• Gate logic. REQUIREMENTS is now test.yml (package-contract) plus production-qualification.yml at the exact SHA. A run counts only if every job in EXPECTED_QUALIFICATION_JOBS succeeded, and that set now includes the hosted Linux job. A skipped or missing Linux job fails closed (covered by test_release_gate_requires_the_hosted_linux_xvfb_lane). The release never queries live-qualification.yml (test_release_requirements_never_include_a_self_hosted_workflow, and the dispatch test asserts on that URL).
• Same tests as the old lane. Both test files run with no -k filter, 8 tests per trial and 3 trials, with no skips allowed. test_hosted_linux_release_lane_runs_input_and_structural_trials pins this.
• Thresholds. None are loosened. The only loosened check is --minimum-monitors 1, because Xvfb reports a single monitor to MSS.
• Scope. The PR body, the workflow headers, DESIGN.md, WINDOW_CAPTURE.md and the CHANGELOG all say the gate no longer proves multi-monitor capture or real input hardware.
• Known risk. production-qualification.yml can only be dispatched on main, so the new job's first GitHub-hosted run happens after merge. I ran the exact YAML step locally (3/3 trials). I'll dispatch qualification at the merge SHA before re-running the release.

Signed-off-by: Richard Abrich richard.abrich@mldsai.com