ci: diagnose the launch-smoke flake, pin artifact retention, gate demo GIFs (#247, #258)

[jeiel85]

Three items from #247 and two from #258, picked up during a sweep over the open issue backlog.

launch-smoke is not the flake it was filed as

Three issues describe this job as an unexamined flake. The failure on PR #274 gave it a name, and it is not Broken pipe:

Performing Streamed Install
Success
Starting: Intent { cmp=com.markleaf.notes/.MainActivity }
Error type 3
Error: Activity class {com.markleaf.notes/com.markleaf.notes.MainActivity} does not exist.

adb install returns when the install session commits — which is before the package manager has the app indexed. Starting the activity in that window fails exactly as though the build were broken. On a slow CI emulator that window is seconds wide; locally it is nothing, which is why the job reads as random.

The script now waits for pm path to answer before starting the activity (up to a minute) and fails with a distinct message if it never does. That is a wait on the real signal rather than a sleep.

The job stays continue-on-error — one green run is not evidence, and #252 already asked for a few more data points before promoting it. But the flake now has a diagnosis instead of a reputation.

Artifact retention

roborazzi-diff-images, lint-report-release and roborazzi-goldens-linux were the three still on the 90-day default. They now carry explicit values with the reasoning next to them:

• 7 days for the two failure-only reports — read while the failure is fresh, or not at all.
• 30 days for the recorded goldens — downloaded and committed by hand, so they have to survive the gap between dispatching the run and getting round to the commit.

Demo GIFs

verify-landing-versions.ps1 now asserts each of the twelve public surfaces references its own markleaf-tablet-<lang>.gif, and that the file exists. README.ko.md pointing at the English clip renders fine and would have shipped silently — the same class of gap this script already closes for version strings.

It lives here rather than in landing-check.yml on purpose: this script runs in the required build check, so it covers the six READMEs too, which the docs-only workflow does not.

docs/assets/README.md records the two remaining #258 items — which version each clip was recorded on, and why markleaf-preview-720/1280.webp are kept despite being referenced nowhere (their GitHub Pages URLs may be linked from outside the repository; ~93 KB total, deleting them breaks those links for no benefit).

Verification

bash -n on the script, YAML parse on the workflow, and the new check exercised both ways: 12/12 surfaces pass as the tree stands, and pointing README.ko.md at the English clip fails it with 기대 markleaf-tablet-ko.gif, 실제 -> markleaf-tablet-en.gif and exit 2.

The install-window fix itself can only be confirmed by the job running on this PR.

Still open on #247: the Release asset list has no CI gate, and gh release create arguments are still only exercised on a tag push.

🤖 Generated with Claude Code