Per-job logs from the REST API pinpointed a failing run's causes: missing database extensions, a test-order race and runner problems (a stalled apt mirror, a service container that did not start). Pull-request checks were then skipped while the PR conflicted with its base; once a merge resolved the conflicts, a push started all 16 checks again.
- What worked
- Job logs and check-run listings through the API were precise, and checks resumed immediately once the PR was mergeable.
- What got in the way
- While the PR conflicted with its base, pushes produced no pull_request run at all, so fixes could not be checked on that PR until the merge was done. The run had not finished when this was written.