Handoff: Finish CI verification for the UrlProtocol admission follow-up; leave it unmerged

Written: 2026-10-20 23:35 UTC.

RE-VERIFY: This is a write-time snapshot. Re-check https://github.com/CodexCoder21Organization/UrlProtocol/pull/749 with gh pr view 749 --repo CodexCoder21Organization/UrlProtocol --json state,isDraft,headRefOid,mergeStateStatus,statusCheckRollup,mergedAt before acting. Fresh checks override this document.

Design decisions in force

The task on 2026-10-20 was to open ONE non-draft follow-up PR carrying the four admission fixes. It stated: “Do not merge, enqueue or mark anything else.” The retention direction was verbatim: “Keep policy, rely on other fixes” (keep up-to-1000 identity retention). “No timeout increases, no iteration reductions, no disabled tests, no retries/sleeps added.” “Before any push, check the target branch is yours and no one else has pushed to it.” “Disk check (df -h /) before every build; stop under 8 GB.” All commits must end Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>; the PR description ends 🤖 Generated with [Claude Code](https://claude.com/claude-code).

This session stopped within the task's 150-minute wall limit while full-suite CI was still running. There is no authorization to merge, queue, deploy, rerequest healthy runs, or open a second PR. Continue verification of the existing PR only. Any change of that scope needs owner direction.

Mission summary

Original motivation: “We need to investigate why https://my.externalbrain.wasmserver.com/ is failing to load and fix the underlying problem.” https://github.com/CodexCoder21Organization/UrlProtocol/pull/725 merged the bounded worker-lane change, but four independent-review admission fixes remained on https://github.com/CodexCoder21Organization/UrlProtocol/commit/d8af97937202bdcc65332be7a82e6f15397da3f2. The requested follow-up is now OPEN and non-draft at https://github.com/CodexCoder21Organization/UrlProtocol/pull/749. Remaining work is required-check verification; do not merge it.

What was found and done

1. Old shared retained-byte refusal recorded a sender SEVERE interaction even when the process's own retained allowance was full. The carried fix scores only the sender's own 64-pending queue overflow and emits a typed, coalesced AnnouncementAdmissionRefusedEffect for local pressure. Queue-full samples retain priority in both coalescing orders.
2. Old retained admission refused any legal announcement whose charge exceeded the whole pool. The carried fix permits exclusive use with a capped retained charge, while the independent wire/parse limits remain in force.
3. Old compact service/relay withdrawals could be dropped on announcement count/byte pressure. The carried fix keeps their parser ownership in the same ordered lane through callback/report completion or host drain.
4. Binary v1.1 withdrawals also keep JSON+attachment charges and detach the physical-root frame turn before handoff, allowing later same-root binary RPC when parser capacity remains. Parser capacity provides backpressure; withdrawal work is not unbounded.
5. Product files exactly match the reviewed admission head. Added one public real-host inverse-order regression and migrated three existing fixtures: bounded refusal reporting now observes typed warnings with zero SEVERE for shared pressure; two withdrawal/root-close tests assert exact held charge then zero after host drain. All old root-close, registry-publication and bounded-report checks remain. No identity-retention change.
6. Read README, scripts/test.bash, full TESTING and PHILOSOPHY guidance. Independent test-coverage and implementation reviews passed, including the extra inverse-order test and fixture migrations. Reviews and findings are saved remotely.
7. Baseline at main's task-start commit, plus warning type declaration only, failed all seven new tests for expected behavior: 21/21 native built-in attempts failed. No admission fix was in that baseline. Initial focused fixed run passed14/14 first attempt; the three migrated fixtures plus a dependency-preparation-blocked stress test passed4/4 first attempt on the rebased final head.
8. Final exact pushed head 9c7a620bd77522776c9a4ce3dd9fb20f6174e3d9 passed 367/367 tests FIRST attempt, across all364 selected files;367 attempts,0failures,0attempt-history records. Every filename containing725/Announcement/Withdrawal/Admission/Gossip/Retained plus all new tests is represented in the native XML. All seven new tests passed first attempt. No test timeout/count/retry policy changed.
9. Whole-suite attempts on the earlier candidate did not complete: remote provisioning never produced a run ID/test progress and the owned attempt was stopped; local full runner exited137 before results, cause unconfirmed. The task explicitly authorized the364-file fallback. Both fallback runs emitted one nonfatal terminal-memory warning beside unchanged testRelayDoesNotReplayStaleCachedAnnouncementToLateRegistrant: output followed the terminal memory record. Final native records prove that test still passed first attempt. Source inspection shows unawaited host stop and unstopped RelayService in that unchanged test, consistent with late output; exact attribution is unproved. No unrelated cleanup or warning suppression was added.
10. Fresh23:34UTC snapshot: required bld-build SUCCESS, required bld-all-tests IN_PROGRESS; OPEN, non-draft, unmerged, no merge queue entry or auto-merge request. Full-suite job: https://github.com/CodexCoder21Organization/UrlProtocol/actions/runs/39514426218/job/118943045350 . Build job: https://github.com/CodexCoder21Organization/UrlProtocol/actions/runs/39514426215/job/118943045206. The watcher also reported a zero-run custom kotlin-build-ci-test suite. Its sampled base suite emits optional kotlin.build (remote) / kotlin.build (kompile-remote-build) checks; required checks are separate bld-all-tests / bld-build GitHub Actions jobs, already dispatched. No rerequest was performed. Do not infer required dispatch loss from the optional suite alarm.

Relevant PRs and refs

No code or artifact was deployed or published by this effort; executed actions were source/evidence pushes, one PR creation, read-only CI checks and local builds/tests. Source checkout is clean with no stashes; all intentional source changes and needed evidence are remote. Build outputs/caches were deliberately excluded. Only owned ~/.aibuildcaches/*jobs_r80* caches were deleted; no own tests/watchers remain active at handoff.

Next steps

1. Re-verify current PR state/head/checks first, at most once per60seconds. If the PR changed or merged, stop and reconcile with owner scope before any push. Do not create another PR automatically.
2. Run the plain watch-only-to-green command: coursier launch buildwatchman:build-watchman:0.0.21 -r https://kotlin.directory -- --repo CodexCoder21Organization/UrlProtocol --pr 749 --idle-timeout-minutes 15. Track it as a bounded background task and read its terminal exit; no --to-merged, enqueue or merge action.
3. If required CI fails, inspect complete native attempt history and logs. Reproduce the actual failure locally through its narrow public path before changing code. Do not blindly requeue, increase timeouts, reduce iterations or disable tests. Check every interface fixture if a shared call changes. Required whole-suite CI may cover cases beyond the locally authorized fallback; that is not a reason to weaken them.
4. Before any build, fetch/rebase current main and check disk>=8GB. Use an isolated checkout under the session's authorized workspace. Before updating this same PR, verify OPEN state, ownership and expected remote head, then push with an exact lease. Any new commit needs the required coauthor trailer, appropriate native checks and fresh independent reviews. Keep the PR non-draft and unmerged.
5. Once both required checks are green, record their native state and report completion of this verification handoff. The explicit request still ends before merge; request no merge unless the owner changes the task.

Operational knowledge

Evidence directory has r80-final-selection.xml, .log, -audit.json, r80-fallback-selection.txt, r80-final-head.txt, initial failed fallback records, both red controls, targeted green records, r80-review.md, r80-findings.md, and CI watch/snapshot records. Download the exact selection from that branch when reproducing the fallback. Native XML, not progress text, is the authority for pass/first-attempt counts. Runner default native attempt histories were not changed. A failed POM preparation download (Broken pipe) was repaired with an explicit successful coursier fetch kompile:build-kotlin-jvm:0.0.33 -r https://kotlin.directory --classpath before its first successful body run.

The report-challenge CLI was not invoked because it automatically creates/merges another PR, contrary to this lane's one-PR/no-merge rule; challenge details are preserved on the evidence branch. Current Handoff skill uses the central service directly with no PR; disregard the legacy merge-queue section still present in the format README.