Handoff: Finish ContainerNursery's EventLog bump when a compatible upstream release is available

Written 2026-10-25, L105.

RE-VERIFY: This is a write-time snapshot. Recheck https://github.com/CodexCoder21Organization/ContainerNursery/pull/749 with gh pr view <URL> --json state,isDraft,headRefOid,statusCheckRollup,mergedAt; read current kotlin.directory metadata and artifact listing before choosing a dependency. Recheck upstream PRs and their source composition before publishing anything.

Design decisions in force

The original 2026-10-25 request authorizes a consumer bump to an already-published upstream fix. Its requirements remain:

Rule: fix upstream bugs upstream; bumping the consumer to a published release that already fixes the bug is the approved path. No downstream workaround.

No reflection, no mocks, no timeout increases, no weakening assertions, no downstream workaround. Do the work yourself; no other model sessions. Wait loops need a hard deadline and must be killed before you finish.

Do NOT merge or enqueue.

The required full-suite command is flock /tmp/claude-501/hd22-suite.lock scripts/test.bash --local --test . --log <file>, with first-attempt counts. The PR description must retain the user's opening quoted in the current draft. No production deployment was requested or performed. Extending this dependency-only task to integrate and publish a new upstream release needs a scope decision; do not assume a new publication is already authorized by this brief.

Mission and current result

ContainerNursery's pinned LocalFileSystemEventLog 0.0.2 catches JDBC exceptions from close(). Consequently EventStore.close returns normally and all shutdown waiters observe success although the SQLite connection is still open. The requested change imports the supplied public shutdown reproducer and bumps to the newest published implementation already fixing close propagation.

The draft PR is concrete and rebased onto the now-merged shutdown prerequisite. Implementation 0.0.8 fixes the supplied close error, but it deterministically fails three unchanged retention-work guards. A green dependency-only bump is blocked on a compatible published upstream release, not on permission to change a consumer assertion.

Findings and completed work

1. Fresh clones under the lane's workspace/: ContainerNursery, community.kotlin.eventlog.localfilesystem and community.kotlin.eventlog. Read repository README/AGENTS where present, Testing Architecture and Engineering Philosophy. No delegates or other model sessions.
2. Mainline upstream fix: https://github.com/CodexCoder21Organization/community.kotlin.eventlog.localfilesystem/commit/4dc9e1c178e8c5b134032142bf1ee9d24bb5ae32. Original feature-branch fix: https://github.com/CodexCoder21Organization/community.kotlin.eventlog.localfilesystem/commit/7ba17b3aab01c12050d7913761f3ff44d0499dd1. Both remove the catch around connection.close; close now unlocks and rethrows the original exception. Static disassembly of actual old/new JARs confirms the difference, without reflection.
3. kotlin.directory metadata AND artifact directory list versions only through 0.0.8. Published source is https://github.com/CodexCoder21Organization/community.kotlin.eventlog.localfilesystem/commit/c9b094b1e0d1cc38a6c29a66307fce32cf80d245, recorded by https://github.com/CodexCoder21Organization/community.kotlin.eventlog.localfilesystem/blob/debcd94c1b617f05f9bd1d00d28e3f202de903b9/findings.md. Main still predates these status-history changes. A new main-based release alone would lose them.
4. Imported byte-identical test from https://github.com/CodexCoder21Organization/ContainerNursery/commit/52d5dc545a3da6aca203eeaa7ae782c9386fb6f2: tests/testL103_concurrentEventLogCloseFailureMustReachShutdown.kts. Test SHA-256 dae5d9243da5628bfc3830766437fbab11cf7a6c71f6e011caf7c5da18a0fa9d. A public JDBC Driver wrapper delegates to real SQLite, holds system_metrics.db close and throws before closing its connection. Nine concurrent shutdowns must receive the same SQLException, as must eight later shutdowns; separate/store-free nurseries still succeed. Checks one failed close, daemon worker exit, actual connection still open, and finally closes all real connections/restores drivers.
5. Baseline .2: FAIL 3/3 native attempts, including attempt one, at Every waiting shutdown must report the metrics EventLog close failure; observed=[null, null, null, null, null, null, null, null, null]. Expected <9>, actual <0>. New implementation .8 + API .4: PASS 1/1 attempt one, zero retry rows. Rebased final head: same reproducer PASS 1/1 attempt one again.
6. Every implementation pin .2 -> .8 and API pin .2 -> .4, including build.kts, build.gradle.kts and 15 test annotation pairs. API .4 is required for TruncatableEventLog. Automated byte audit verifies these 17 files have only the two version substitutions, preserving all other bytes, including binary-containing test scripts. No consumer implementation or assertion change. Existing shutdown README already states retained-close-error behavior.
7. Full suite ran once under the requested lock, after acquiring it and fetching/rebasing. Tested head: https://github.com/CodexCoder21Organization/ContainerNursery/commit/10d59c56f658e9e120f9f6143a50666361592d30. Results: 1,090/1,093 attempt-one passes; three first-attempt failures, three final failures, three retry histories. All three failures are deterministic 3/3 native attempts: testEventStore_systemMetricsRawRetentionScale reads 199 rows for 154 buckets; MinuteRetentionScale 247 for 202; TenMinuteRetentionScale 307 for 262. All fail at line 199's unchanged rowsRead <= expectedRows + 32. Close reproducer passes attempt one in this run. Do not present the suite as green or retry it unchanged.
8. Temporary query instrumentation preserved every assertion and proved exactly 45 non-bucket rows: auto_vacuum=3, freelist_count=6, table_info(events)=6, table_info(system_metrics_cursor)=6, MAX(sequence)=9, MIN(sequence)=9, system_metrics_cursor=1, projection_cursors=2, reclamation_pending=3. The 154 raw bucket rows are correct; this is added fixed metadata work, not history replay or timing. Temporary diagnostic removed from the code branch, patch archived. Unchanged raw-retention test on old .2 pins passes 1/1 attempt one, ruling out an already-failing base guard.
9. Upstream https://github.com/CodexCoder21Organization/community.kotlin.eventlog.localfilesystem/pull/16 is OPEN at https://github.com/CodexCoder21Organization/community.kotlin.eventlog.localfilesystem/commit/69ec1e5f4193c72208e8630da4061b658892bb3c. Commit https://github.com/CodexCoder21Organization/community.kotlin.eventlog.localfilesystem/commit/8d0ebde2919cd11c7798e24c2b45f8d3f4c6b2a5 consolidates the repeated startup and MIN/MAX reads. That PR independently records the same 45-row consumer failure, fixed local consumer smoke passes and 68/68 upstream native passes. Those are upstream author's results, not this lane's qualification. It explicitly says version 0.0.11 is unpublished, publication prohibited by its brief, and this main-based source must be integrated with the status-history changes in published .8 before release. Do not publish its branch by itself. Published status-history changes are tracked in https://github.com/CodexCoder21Organization/community.kotlin.eventlog.localfilesystem/pull/15.
10. Shutdown prerequisite https://github.com/CodexCoder21Organization/ContainerNursery/pull/745 MERGED at 14:36:27 UTC. Rebased only the two lane commits onto origin/main a9d863cc. Main's only intervening tree change relative to our full-suite head is its independently updated testHttpsFacade_largeChunksAndTrailersUseBoundedTlsRecords.kts (119 additions/10 removals). That test passes 1/1 attempt one at the rebased head, as does close reproduction. Full-suite counts remain attributed to the original head; no second full suite or green full-suite claim. Final branch head is e65dcb394c224852ba643b39a5bfbf694665a1c7; remote verified equal. PR scope against main is only the imported test and 17 dependency-pair substitutions.
11. Separate same-session test-comprehensiveness and code-review passes completed; invariant table and byte audits in archived findings. All diagnostic changes removed; no notes, logs, jars or build artifacts committed to the PR.

Other behavior changes reviewed in .2 -> .8

Commit-ordered batching (up to 256 appends) and durable shared sequence allocation; fair database locking and transactional reads; one-time existing-DB incremental-auto-vacuum conversion using VACUUM, metadata table and busy_timeout=30000; bounded 512-page background reclamation; optional truncation and sequence reads with resumedPastTruncation; synchronized iterators/infinite-wait overflow handling; explicit pending-append drain and drain/interruption errors on close. IDs become opaque instance-scoped receipts; successful status uses a watermark, failed statuses retain a 4,096-completion window and expire the preceding prefix to unknown. ContainerNursery ignores append receipts, uses fetchBatch with persisted cursors and never calls truncation. It therefore does not depend on old receipt lifetime/format or delete history from this bump. Full details and upstream diff links are in the PR and archived release audit.

Artifact verification

Downloaded bytes equal their repository-listed .jar.sha256; resolver-cache implementation JAR checked too:

• Implementation .8, 41,327 bytes: 45f69653db9ed54fe5ad5b0cded5616b20837115c4bd367135a8a5f6d73fe2ad.
• API .4, 12,736 bytes: c6efb57083230a7ef865e85c7ebff15571fc8b4fda64d3fd3d94bc8369c9d557.
• Baseline implementation .2: 37708f0be63f2ba39ab9c24805923b6c0585d1e54ef504ef22dedc52fc007130.

Relevant PRs / refs

Upstream clones were read-only, no local changes or stashes. The temporary diagnostic is an archived patch, not pushed product code. Full native reports, logs, counting script, diagnostic patch and findings are in the evidence link below. Downloaded JARs and build/cache outputs are deliberately not uploaded or committed; redownload published artifacts and verify listed hashes instead. No source changes or partial fixes exist only locally.

Deployed/published state

This lane deployed, merged, enqueued and published nothing. The .8 upstream artifact predates this task and contains c9b094b changes not on upstream main; this matters when composing a future release. No claim is made about live production host state.

CI and tooling snapshot

Initial-head canonical CI https://github.com/CodexCoder21Organization/ContainerNursery/runs/121197469807 failed before source upload or tests. Upload request c4a0ebf3-f525-48b7-bc51-ca569ffcf596 made five createUploadSession attempts, zero sessions, zero chunks. Output reports two queued duration events for b95d4f17 could not publish under /root/buildtest-data: java.io.IOException: No space left on device. New publications are refused. This is a separate shared CI blocker, not evidence of current free space or disk-growth mechanism. Full stack is in archived canonical-ci-diagnostic.txt and the linked check. Rebase legitimately creates a new head/check; no manual rerun or server operation was performed. Recheck final-head CI live; green is not claimed.

Runner .124 also emits TerminalMemoryWarningEffect when stdout arrives after its terminal memory record. Native full-suite XML gives the actual failures/counts; this metadata warning is not their cause. No runner workaround. report-challenge CLI was not run because it automatically enqueues/merges a separate PR, conflicting with this task's prohibition. Installed gh pr edit fails on deprecated Projects(classic) GraphQL; use REST PATCH with JSON input file and read back the body. No shell interpolation of multiline PR text.

Next steps

1. Re-verify current PR/main/upstream release and CI states. Read this handoff and archived findings before editing.
2. Resolve the scope/publication dependency: use an already-published integrated upstream release if one now exists. Otherwise obtain direction to extend the original consumer-only task to integrating and publishing upstream changes. Preserve .8's status-history fixes while adding the startup/read-wait fixes; never release the main-based PR branch alone. No consumer workaround, assertion weakening or timeout increase.
3. Audit the selected release diff and API dependency; consistently update all implementation/API pins. Byte-verify downloaded artifacts against listings.
4. Keep the supplied close reproducer unchanged. Require its before/after evidence and all three unchanged retention guards to pass attempt one. Run the full suite under the shared lock for the new artifact, record first-attempt counts, and perform separate source/test review passes.
5. Before pushing verify the draft remains OPEN, fetch/rebase main, verify the exact tested head, push with lease if necessary and update the PR opening/versions/counts. Watch canonical remote CI and required checks via bounded build-watchman. Do not rerun unchanged failures. Do not merge, enqueue or deploy under this brief.

Evidence

Final native reports, logs, query diagnostic, late rebased-head checks and findings: https://filedrop.wasmserver.com/drop/d377074942 Archive SHA-256 f80a11ee9159827d56fbaa45c9f4cd8dd708712428949e0321b4a8866006ba1d; one-year retention requested. Earlier evidence snapshot: https://filedrop.wasmserver.com/drop/0d471433f0.

Write-time final snapshot 2026-10-25 14:44 UTC: draft is OPEN against main at e65dcb394c224852ba643b39a5bfbf694665a1c7; final-head canonical remote check is FAILURE, both GitHub Actions build/release jobs and informational remote-build check are still in progress. Initial diagnostic above establishes the shared pre-upload publication blocker; inspect final-head output rather than assuming an identical request ID. Both task-owned watchers have ended/stopped; no timers or test processes are intentionally left running. All code is pushed. No approval to publish or merge has been inferred.

Brief audit: fresh clone/branch/README/AGENTS and prerequisite rebase DONE; latest published version/upstream commit/behavior diff DONE; imported baseline red and upgraded green plus all consistent pins/SHA verification DONE; requested single full suite and attempt-one counts DONE (three deterministic failures, not a passing suite); pushed draft PR with required opening and all requested evidence DONE. A green dependency-only upgrade on the latest published artifact is REFUTED by the unchanged retention guards and exact metadata counts, requiring an integrated upstream release before this draft can become ready.

Final-head CI diagnostic verified 14:45 UTC: https://github.com/CodexCoder21Organization/ContainerNursery/runs/121199221098 failed before upload with request e27ba169-5463-46ff-a8ad-246004f65593, five createUploadSession attempts, zero sessions/chunks, same b95d4f17 publication refusal under /root/buildtest-data. Both task-owned watch processes have exited (prerequisite success, own watch stopped after terminal red and release blocker established). No automated watcher, local test or timer remains.