[rmcdaniel]

This fixes two CI flakes that failed #851's merge-queue run (run 40824302270) and earlier PR runs. Under the CI-first rule they come before more feature PRs.

Composer floor image build

scripts/ci/brotli-safe-composer.sh builds the pinned Composer image with Alpine's fixed brotli-libs (CVE-2026-111497). The build failed now and then with apk add --no-cache --upgrade "brotli-libs>=1.2.1-r0" exiting 5 while it fetched the package index. That fails every job that installs Composer packages: for example #851's MariaDB restarts lane, and #850's corrected-markers lane on d85908cb.

The build is now retried twice, 10 and 20 seconds apart, before the failure is reported. The image is still checked on every use, as before. A new composer_trust_test.py case runs the script with a fake docker whose build fails twice, then three times.

Installed scheduler wait

schedule-probe.php gave published PHP's installed schedule:evaluate role ten seconds to fire the first occurrence of a 2-second schedule. Each pass of the role's loop boots the application twice (repair pass, then schedule evaluation), and on a loaded runner that can take more than ten seconds. #851's merge-queue corpus failed there: the excerpt read schedule-fixed-rate-one-occurrence (runner-blocked): ... did not fire the original occurrence within ten seconds. It now waits 25 seconds, inside the runner's 45-second probe budget. The fixture's max_runs: 1 keeps a longer wait from firing more occurrences.

Verification (local)

• python3 -m unittest tests/Unit/Support/composer_trust_test.py: 20 run, 6 skipped (they need Docker), all others pass.
• The new retry test fails against the previous script.
• sh -n on the script passes.

🤖 Generated with Claude Code

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[rmcdaniel]

Reviewed exact 1dae4ff. The Composer floor build makes at most three attempts with 10/20-second delays, exits nonzero on exhaustion, and retains the pinned-base label and installed decoder checks on every use. The new test runs the actual shell script with a Docker stand-in and checks both recovery on attempt three and failure after three attempts. The local 20-test / six-skip result remains author-attributed; I inspected source without rerunning it.

The scheduler change extends only the observation deadline to 25 seconds inside the unchanged 45-second probe budget. The fixture still uses the original committed two-second occurrence, max_runs=1, its original workflow/audit links and two later real interval checks for no extra occurrence. The probe does not move the scheduled deadline or write target rows. This does not establish a scheduler-latency guarantee.

No targeted source finding from this review. Normal exact-head hosted and protected cumulative qualification remain required; prior failed runs remain failed evidence.

Exact-head hosted qualification at 1dae4ff is now terminal: at 06:36 UTC on October 29, the primary rollup reports 56 successful checks, one expected skip and no open or failed check. Protected cumulative qualification and merge remain required.

Protected qualification is now verified: cumulative e6c154b passed Shared Server fixtures 40831096377 at 07:06:31 UTC, PHPUnit Feature 40831096518 at 06:59:40 and Public Boundary 40831096468. PR #852 merged at 07:06:56 UTC as that same commit. This is source/CI delivery, without a new published Server image or installed-runtime claim.