Handoff: Publish UrlProtocol with the TCP stop fix and finish the ContainerNursery bump

Written 2026-10-20 16:47 UTC.

RE-VERIFY: This is a write-time snapshot. Recheck GitHub PR/branch state, the current Maven artifact listing, and direct candidate POM/JAR URLs before acting. Use gh pr view <full-URL> --json state,mergedAt,mergeCommit,statusCheckRollup, git ls-remote, and direct kotlin.directory GETs. The selected version may have been published by another session since this snapshot.

Design decisions in force

The assigned task is "Publish UrlProtocol with the TcpServer.stop() fix and bump ContainerNursery to it" (2026-10-20).

The user's constraints, verbatim where decisive:

• "Build locally (--local, -Xmx2g), not --remote, then upload."
• "Coursier: use /home/u/bin/cs (other copies are broken)."
• "Do NOT edit tests/testUrlFacadeRpc_reacquiresContainerOnDeadPort.kts — it stays unchanged and is the proof."
• "Never raise a timeout, never lower an iteration count, never weaken/delete an assertion or test, no reflection, no mocks."
• "Never re-run CI to get green; any flaky test you hit gets root-caused (fix-flakey-test skill), not re-run past."
• "Do NOT enqueue or merge, and do not touch the merge queue — the supervisor may enqueue PRs itself; leave its actions alone. Do not close PR 672 (the supervisor will)." The referenced draft is https://github.com/CodexCoder21Organization/ContainerNursery/pull/672.
• Full CN local suite once under flock -w 1800 /tmp/cn-fullsuite.lock; check df -h / first, clean only task-owned build outputs if needed, and report a disk failure separately from test results.
• CN PR ends with 🤖 Generated with [Claude Code](https://claude.com/claude-code); commit trailer Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>.
• Final requested fields: published version + SHA256, PR URL, head SHA, before/after first-attempt counts, full-suite counts, bld-build verdict.

These constraints supersede any general merge/deploy or timeout-change authority. No source implementation or test assertion was changed in this lane.

Mission and why

The original user asked on 2026-10-19: "The ContainerNursery service should keep track of all recent jars deployed for each service, along with the date/time it was uploaded and the date/time that each of the jars was the jar actively serving the site and make it easy to roll to a different jar (eg. roll back a bad change), etc." and "This may involve updating the CLI/api too. Merge your changes, ensure test coverage is comprehensive."

While validating that work, the unchanged CN integration test containernursery.testUrlFacadeRpc_reacquiresContainerOnDeadPort intermittently returned RPC error: INTERNAL_ERROR - Stream closed on main. The owning UrlProtocol fix is already merged. This lane must publish it and verify the consumer upgrade with that original test, rather than rewrite the test.

What was found and done

1. Fresh clones were created under the requested lane's workspace/UrlProtocol and workspace/ContainerNursery. Read each README, existing root AGENTS, build/test scripts, the publish-maven-artifact skill at the requested .claude path, Testing Architecture and Engineering Philosophy documents, and the linked dead-port investigation. UP has no root AGENTS; CN requires isolated tests and no mocks.
2. UP main at source inspection was 7b837af3f466c65314ffba642cf1011653d39671. https://github.com/CodexCoder21Organization/UrlProtocol/pull/741 merged at 2026-10-20T15:47:46Z. It changes stop to wait for accept completion, preserving self/reentrant stop and failed-close retry behavior. Its bld-build and bld-all-tests were successful at the fresh check; consult the stored status JSON and recheck live.
3. Published foundation.url:protocol:0.0.1425 was downloaded and inspected with javap -c -p. Its stop method has no accept-completion wait; the other Thread.join in that class belongs to reader cleanup. The output is preserved as tcpserver-1425.javap on the evidence branch. Direct candidate 0.0.1426 POM probes returned404, last at 16:41:55 UTC.
4. Prepared and pushed the UP coordinate-only branch below. It changes main's1425 coordinate to1426 and adds a dated history comment linking the merged fix. Its src tree hash is exactly main's 25effe05111482da4f9928980fdb3826e4ef9f99. No implementation code differs from merged main.
5. Neither local UP build completed. Both used --local, JAVA_TOOL_OPTIONS=-Xmx2g, kompile-cli0.0.124 and the existing20min rule deadline. First exit1: BuildScriptTimeoutException, elapsed1200197ms/deadline1196324ms; second exit1: elapsed1195862ms/deadline1191392ms. Both children were terminated by the builder. No JAR output was produced and nothing was uploaded. These are compilation failures, never test results.
6. Compiler samples progressed through frontend, IR and bytecode generation in the first run; the second reached frontend type analysis. At one first-run sample the main thread had173 CPU seconds in1114s elapsed and17695 major page faults. GC accounted for12.2s and no full collection; timeout heap was~392MiB live of~1090MiB committed, max2GiB. Paging was observed early. Later memory pressure subsided but CPU pressure remained: container CPU PSI full avg60~56.7%, host load~64 on8 available CPUs, and a second-run main sample had29 CPU seconds in335s elapsed. These are measurements of scheduling delay, not a complete diagnosis of every compiler cost or proof that hardware must change. Do not raise limits or keep restarting the same build unchanged; resolve/recheck the local compile scheduling condition first. No other lane's work was stopped.
7. The initially warming CN target runner was cancelled while its main sources were compiling, before any test executed, so it supplies no attempt count. This avoided running our two large source compiles concurrently. All task-owned build and measurement processes have now ended.
8. Ran the original linked TransportEvidence.java against published1420: ten fixed independent JVM batches,100 real lifecycles each, four concurrent lifecycle callers, -Xmx128m. All10 batches failed the post-stop-no-connect contract. Aggregate:1000 trials,627 connection refusals,370 successful post-stop connections,289 exact Stream closed replies,3 separate reset-before-open outcomes. Exact Stream closed occurred in every batch. This is adjunct transport evidence, not the requested CN runner comparison and not fixed-version evidence. Raw logs and summary JSON are preserved remotely.
9. Independent test and source/API review is complete, no blocking finding. The eight upstream regression tests cover controlled accept completion, concurrent/reentrant stop, failure/retry, late accepts and restart isolation. UP stop retains its JVM descriptor and this fix adds no dependency. Keep CN's resolver1320 unless runtime evidence requires otherwise. The existing CN ABI test separately pins protocol1420 twice; both literals must move with build.kts. Its scanner covers method calls, skips constructors/fields, and is not a complete ABI proof. Full review is preserved remotely.
10. CN original oracle remains byte-for-byte unchanged, SHA256 26ee32bcc5610771739a11085e488a6d482e33427f4e2717c06208c1c646e856. Prepared an apply-checked cn-bump.patch changing only build.kts pin/history comment and the ABI test's two version literals; not applied. Prepared the PR body with exact why, upstream link, supersedes link and footer; not opened.
11. Prepared OracleBatch.kts.txt: the original test source followed by199 distinct wrapper functions, each invoking the original test function, giving200 named normal-runner cases. Each call creates fresh fixtures. Report first-attempt XML history, not retry-adjusted final success and not claimed statistical independence. This batch was not run. It is stored as .kts.txt so it cannot accidentally enter the ordinary full suite.

Relevant PRs / refs and remote checkpoints

Upstream fix: https://github.com/CodexCoder21Organization/UrlProtocol/pull/741 — merged. Superseded CN test-rewrite draft: https://github.com/CodexCoder21Organization/ContainerNursery/pull/672 — freshly verified OPEN and draft, left untouched.

CN base used here is 44da3cacb6849c7fa297887ef84c6de0e0801f3b. Local fix/bump-urlprotocol-tcpstop still points to that base with no committed bump. There is no remote functional CN bump branch or PR from this lane. The evidence branch is separate deliberately; do not carry its diagnostic files into the eventual bump PR.

Published or deployed but not merged

None. Both local builds failed before producing a JAR. No upload/deploy command was issued. Direct1426 POM still404 at the last probe. The already-published1425 artifact predates this effort and lacks this stop fix. Do not represent1426 as published or fabricate a digest.

Remaining steps, in order

1. Re-verify current source, PR and Maven state first. Read current required skills/instructions and fresh READMEs. Obtain a successful local main-source build inside the existing deadline; investigate the compile scheduling/CPU delay as needed. No timeout increase, remote build, or CI debugging substitute.
2. Check1426 is still free. If source's candidate coordinate was taken, choose the next free one and update the small UP version PR accordingly. The supervisor may enqueue/merge; this lane may not. Rebase before any build and immediately before PR creation/update.
3. Build locally from main at/after7b837af3 using -Xmx2g, publish with the Maven publisher0.0.6+ through --artifact-dir ... --api-url https://api.kotlin.directory/upload --force. Publication was explicitly authorized. Use /home/u/bin/cs. Verify the downloaded JAR bytecode has acceptLoopCompletion/CompletableFuture.join in stop and record its SHA256, matching built bytes.
4. From fresh latest-main CN, branch fix/bump-urlprotocol-tcpstop. For strict before/after keep the original test untouched. Copy the saved .kts.txt to an untracked diagnostic .kts and run it on the old pin, then the bumped pin with the same runner settings; preserve XML first-attempt counts. summarize-results.py understands <test> plus <attempt-history><attempt number="1">; zero test cases means no test verdict. The stored run-oracle helper assumes the original lane/evidence/workspace layout; adapt paths, not test bodies. No baseline count currently exists.
5. Apply/rebase the prepared patch, changing candidate version if publication selected another one. Keep resolver1320 unless concrete compatibility evidence requires a change. Run the unchanged oracle repeatedly, all relevant URL facade neighbors and the aligned ABI test; then remove the diagnostic .kts and run full CN local suite once under flock -w 1800 /tmp/cn-fullsuite.lock. Check disk first and distinguish preparation/disk failure from test failure. Last disk reading was~4.3GiB free; only task-owned outputs may be cleaned.
6. Commit/push, finish any review gaps, open the CN PR using the saved template plus real before/after and full-suite counts. Footer/trailer and both links are specified above. No PR may claim pending validation passed.
7. Watch required bld-build with watch-build/build-watchman to a verdict, fixing real failures locally first. No enqueue/merge, no queue changes, no closure of the old draft, no production deployment.

Local operations and retained/omitted files

Original lane: /tmp/claude-1000/-code/28a83ccc-c521-474c-b074-68e791f803a6/scratchpad/lanes/publish-urlprotocol. Requested progress file: /tmp/claude-1000/-code/28a83ccc-c521-474c-b074-68e791f803a6/scratchpad/out/publish-urlprotocol-findings.md.

Working Coursier is /home/u/bin/cs. The repository launcher was generated with that Coursier's fetched kompile-cli0.0.124 classpath so broken alternate launchers were not used. CN's bootstrap script verifies a different native Coursier digest, so the planned runner helper invokes the same generated kompile launcher directly with the normal cache/workspace/test/log arguments. No tracked script was edited.

First build command from the lane was JAVA_TOOL_OPTIONS=-Xmx2g workspace/UrlProtocol/scripts/build.bash --local foundation.url.protocol.buildMaven evidence/urlprotocol-maven. Second, from UP after fetch/rebase, was JAVA_TOOL_OPTIONS=-Xmx2g timeout 1500 scripts/build.bash --local foundation.url.protocol.buildMaven ../../evidence/urlprotocol-maven. Internal rule deadlines were unchanged. Full raw errors are on the evidence branch.

A completed reviewer left a finite-but-uncapped javap inspection loop running after its final message; it and all descendants were stopped. No source, test, CI, watcher or inspection loop remains active. No session cron tool was exposed and no recurring background loop was installed. There is no active goal.

Build outputs, compiled .class files, downloaded dependency JARs and classpath launchers are deliberately not pushed; they are reproducible build artifacts. Canonical Testing/Philosophy/README documents are linked in the preserved independent review instead of duplicating them. The useful source, patch, counts, review, failure logs and thread samples are pushed and remote SHAs verified.