Handoff: Finish the first-attempt CI audit for UrlResolver relay verification

Written 2026-10-28 16:31 UTC. RE-VERIFY: this is a snapshot. Check https://github.com/CodexCoder21Organization/UrlResolver/pull/1584 and https://github.com/CodexCoder21Organization/UrlResolver/actions/runs/40711098683 before acting. Poll no more than once every four minutes; use GraphQL when REST is exhausted.

Directions in force

2026-10-28: “Never merge, enqueue, deploy or publish.” “Rebase only on a real conflict; if GitHub reports BEHIND with no conflict, do not rebase.” No raised timeouts or limits, reduced iterations, weaker/removed assertions, reflection, mocks or JVM tuning. Use targeted local tests; do not start or wait on remote buildtest suites. Evidence stays outside the PR diff.

Mission

The owner reported “I feel like https://droplets.wasmserver.com/ is super unreliable ...” and “...super out of date when it should be fast and reliable”. This audit decides whether the resolver relay-route change is READY, without merging or deploying it. Lane CEN1584b has a 90-minute bound from 16:09 UTC.

Findings

1. PR head is d5e2fd6183e2547e89696a393ec08340ac9f4764, OPEN/MERGEABLE at pickup, BEHIND without conflict. No rebase or product change was made. The strengthened relay-only proof and earlier 12/12 local first attempts remain in the original evidence.
2. Original Actions https://github.com/CodexCoder21Organization/UrlResolver/actions/runs/40523067676: 3,054 tests; 3,051 first-attempt passes; three retried; four extra attempts; no final failures. A second XML parser independently confirms these counts and validates ordered histories and final/last-attempt agreement.
3. Main Actions https://github.com/CodexCoder21Organization/UrlResolver/actions/runs/40518327381 at the PR's base 1cbf26353e39eecd3565059df9fd051cf8e03438 contains the exact reply-loss timeout assertion and WITHDRAWAL_NEVER_PROCESSED failures. Main https://github.com/CodexCoder21Organization/UrlResolver/actions/runs/40526682927 repeats withdrawal failure. Those retries are PRE-EXISTING, including the formerly unproven timeout variant. All 12 downloaded main XML files are counted separately, with complete histories.
4. The worker retry is PRE-EXISTING too. Its diagnostic appears before retirement-task completion; an explicit close can still own and return the same cleanup aggregate. The deterministic gate from https://github.com/CodexCoder21Organization/UrlResolver/commit/da10b439 fails 1/1 on PR head and 1/1 on clean main a2b445b451432e129929c7cf7c79ee49d09e2d83, no retries, with the exact aggregate/cause/suppressed chain. Only the function name was changed for the diagnostic probe; its assertions remain. Main already corrects this test via https://github.com/CodexCoder21Organization/UrlResolver/pull/1586 .
5. Read the three tests and exercised code. Four sandbox files and all three old test files are byte-for-byte unchanged from the PR base. Direct gossip adapter/send/batch sections also match. The withdrawal test disables relay and uses the injected-network early return; the worker test creates no resolver network. The relay-shutdown test does use changed relay-candidate helpers, but its exact failure is already in main XML. No PR-caused failure is confirmed.
6. Native head controls passed interrupted-worker 1/1 and withdrawal 1/1. These passes are not used to dismiss flakes. The temporary deterministic probes were removed after saving their source and XML; source checkouts are clean.
7. Fresh Actions is still running at the 16:30 watcher poll, about 92 minutes into a run. Recent completed main runs took 74–101 minutes. Watcher PID 26661 uses plain --pr, default 240-second cadence, and a Python Popen.wait(timeout=3600) deadline from 16:13:45 UTC. Its parent exec session is 35508. The non-gating remote app suite has zero dispatched runs; no rerequest was made.

Refs

No code was merged, enqueued, deployed or published by this lane. No production access was made. Build outputs and repetitive compiler warning logs are excluded; XML keeps complete failure traces.

Next steps

1. Re-verify live state. Continue the tracked plain watcher until its PID exits or deadline fires; do not add a polling loop. Kill every owned watcher before leaving.
2. Run download_fresh.py from the evidence branch when Actions ends, then audit_xml.py fresh-actions. It checks the exact head and collects every XML in test-attempt-records. Read every first-attempt failure and compare its path with main; keep the three earlier PRE-EXISTING proofs.
3. Fix only a confirmed PR-caused failure, test-first on the PR branch after confirming it remains OPEN. Otherwise no product edit is needed.
4. Push fresh XML/counts/findings to wip/CEN1584b-evidence; update the PR description with why, mechanism, validation, first-attempt counts and evidence links. Return READY or NOT READY and update/complete this handoff. Never merge or enqueue.

Running controls

SingleAttemptCli.java/install_single_attempt.py on the evidence branches wrap the public workspace factory with maxAttempts=1 in the ignored launcher. Use scripts/test.bash --local --test <name> --log <outside-repo.xml>. worker-ordering-probe.kts is the deterministic old-expectation probe; copy it temporarily as tests/testCen1584bRetirementOrderingEvidence.kts and remove it after collecting XML. The complete traces are in worker-comparison.json and both worker-ordering XML files. failure-audit.md and path-comparison.json record the read paths and exact unchanged sections. Commit with Claude/noreply@anthropic.com and the brief's co-author footer. The owner override forbids rebasing merely to catch up with main.