Collect three SearchableBucketEmbedded first-attempt reports after runner artifact retention lands

Written: 2026-10-27 11:08:44 UTC by fable-W5-1027.

RE-VERIFY: This is a write-time snapshot. Read PR states with gh pr view, remote heads with git ls-remote, main workflow/scripts and effective rules with gh api, and every test's attempts from downloaded test-attempt-results XML. A final ALL TESTS PASSED summary does not establish first-attempt cleanliness.

Design decisions in force

Owner direction, 2026-10-27: "Use your best judgement, do not ask blocking questions."

Current orchestrator ruling: use GitHub bld-build runner evidence rather than waiting on recovering BuildTest. Collect per-test first-attempt records from at least three main-branch runs where available, plus read the three earlier admitted remote runs if they have executed. Complete only when three inspectable runs have zero first-attempt failures; otherwise fix each demonstrated defect test-first. Deleted evidence is not a blocker: regenerate it. No merge, queue action, approval, production action, timeout increase, test iteration cut or assertion weakening is authorized for this worker. The orchestrator owns merges and CI waits. The full-suite gate is bld-build on the GitHub runner; do not run a full local suite on this shared worker.

Mission and verified findings

The original handoff asked to investigate eight first-attempt failures in remote run 5b93c074 after the recovery change. Historical records were removed by retention. Earlier W1 submitted 00653c5b, b360f7c9, b80e0a89, but their clients expired while the runs were pending with no attempts. Those snapshots are preserved on the W1 evidence branch; W5 did not treat them as current facts or clean results.

W5 claimed successfully and cleared the stale remote-scheduling blocked reason. Main is cf052452ce5b0db549cb176337b7ef7c48376f9f. Recovery and reader-cleanup changes are already merged at https://github.com/CodexCoder21Organization/SearchableBucketEmbedded/pull/21 and https://github.com/CodexCoder21Organization/SearchableBucketEmbedded/pull/22. The separate interruption change https://github.com/CodexCoder21Organization/SearchableBucketEmbedded/pull/23 remains open at 66b7c6e993e5b5a2badfc352f9df6d780b20b32d at the W5 read; it was not modified, closed or adopted by this item.

The requested racing setup already exists: scripts/build.bash and scripts/test.bash pin kompile-cli 0.0.124; .github/workflows/bld-build.yml runs both build and full tests with --local-and-remote, including --test . --log test_log_file.xml. Main's effective required status check is bld-build (integration 15368), not a remote-only check. No ruleset or toolchain upgrade is needed.

Three newest main bld-build runs were read directly:

The mechanism is specific: the workflow generates test_log_file.xml in its disposable runner but never uploads it. Successful suite summaries cannot establish each first attempt passed. W5 opened https://github.com/CodexCoder21Organization/SearchableBucketEmbedded/pull/24 to preserve that existing file with an always-run actions/upload-artifact@v4 step, named test-attempt-results, with 30-day retention. This is eight workflow lines only; all build/test commands, pins, timeouts, assertions and triggers are unchanged. The pinned CLI manager includes attempt-history XML support. Actionlint 1.7.12 and git diff --check passed; an independent static review found no blocking issue. Hosted upload and usable XML remain for CI to verify. No local full suite was run for this CI-only change.

Each earlier remote run was read once through the existing read-only collector. All three calls failed the initial health RPC with CONCURRENT_HANDLER_LIMIT (TcpServer port 46125 at 256 request handlers), before metadata or attempt retrieval. Current run status/counts therefore remain unverified. No run was submitted, retried, attached, cancelled or waited on. This remote failure is not a dependency for obtaining runner evidence.

Relevant PRs / refs

W5 evidence directory: https://github.com/CodexCoder21Organization/SearchableBucketEmbedded/tree/evidence/w5-first-attempt-retention-20261027/handoffs/artifacts/w5-runner-attempts-20261027 . Raw logs retain their original whitespace and hashes. No binary build artifacts, caches or credentials were committed. W5 checkout is clean, with no stash or unpushed commit; both branch heads were checked against git ls-remote. Nothing was deployed or published by this worker.

Next steps

1. Re-verify live state first. Review and land https://github.com/CodexCoder21Organization/SearchableBucketEmbedded/pull/24 through the existing required bld-build gate. Confirm test-attempt-results contains usable test_log_file.xml. No ruleset, deployment or Maven publication step is required.
2. Collect three main-branch bld-build runs with the new workflow: the merge-triggered run and two workflow_dispatch runs are sufficient if all actually execute the full suite. Use gh workflow run bld-build.yml --ref main --repo CodexCoder21Organization/SearchableBucketEmbedded for the two independent fresh runs. The orchestrator owns their waits; do not wait on BuildTest.
3. Download test-attempt-results from each run and parse the XML. Reconcile all selected test identities, inspect each attempt-history/attempt number 1, and retain every first failed output/stacktrace even if a later attempt passes. An absent report, incomplete selection or final summary alone is not a clean run. Preserve the three XML reports on an evidence branch before artifact retention expires.
4. If a first attempt failed, reproduce the named failure reliably before changing code, record red/green/revert-control, fix it at its owning layer, and open one PR per independent defect. Do not infer a root cause from summaries or add retries/timeouts. Search existing PRs before duplicating a fix.
5. When three inspectable main suites have zero failed first attempts, report DONE and complete this handoff. Until then it stays open-claimed for the orchestrator's continuation. There is no current demonstrated flake fix, no deployment, and no publication waiting on this PR.

Operational notes

Use gh api repos/<owner>/<repo>/actions/jobs/<job-id>/logs for raw runner logs: gh run view --log returned empty stdout with exit 0 in this session. W5 job IDs were 121577294395, 121525023564, 118762626408. gh pr edit hit a Projects classic GraphQL error; updating the body through the pull-request REST PATCH succeeded. The report-challenge CLI was not invoked because it automatically merges, which this assignment forbids; CHALLENGE.md preserves the findings for the orchestrator.

Read: github-repos, create-handoff, status-report, kotlin-ci-setup and report-challenge skills; complete handoff-triage README; SearchableBucketEmbedded README; externalbrain github-builders-racing-remote-ci.md; public CLI XML-attempt regression sources and upload-artifact action definition. The independent reviewer also read the supplied Testing Architecture and Engineering Philosophy guidance. No background runner or watcher from W5 remains.