Fix core CI failures, finish UrlProtocol remote checks, and release the final CLI

RE-VERIFY — 2026-10-08 14:18 UTC worker w1 PARTIAL snapshot. Use gh pr view <full URL> --json state,headRefOid,statusCheckRollup and git ls-remote before continuing. The sweep worker cannot merge, enqueue, close, complete, or deploy. The orchestrator owns those decisions. Deploy steps are skipped per the sweep request.

The user asked to work through open handoffs and drive desirable engineering changes to completion, skipping deploy and disk-exhaustion steps. This handoff remains desirable: current main still has the ownerless effect notice and the upload callback-close gap. Resume, poll, archive and reconnect PRs named in the previous body have already merged; no need to repeat them.

Verified work and mechanism

UrlProtocol reports a close/admission record to stderr, then tosses the effect on the callback thread when its creator had no effect repository. The fix makes stderr the full report in that case; real owner-scope delivery and handler/sink failures remain observable. Both real-host tests fail on fresh current main (0/2); the rebased candidate passes 2/2 and the 17 close/admission neighbors pass 17/17. Full local suite passes 2339/2339 on parent 7a10f88926f532051b877b78ae334d2bf095cf9f. Final head 3d62d0de5b4f6ab1fbabd1b3a0f138e9d18a59d4 changes only the test cleanup to attempt both host stops and preserve failures; both regressions pass 2/2 again. This final head is pushed. bld-build and bld-all-tests are SUCCESS; required kotlin.build (remote) is IN_PROGRESS at https://buildtest.kotlin.build/run?id=4ca9e65a. Read-only Fabric projection at14:16 UTC confirms TESTING,2339 total,12 passed,0 failed. WUI cached PENDING is stale. Final-head post-green reviews remain unfinished until remote is green. Never accept an old-head or red required check.

CLI consumes already-published protocol 0.0.569, whose downloaded JAR matches SHA-256 b54681c41749bec201bcbb530a51be66d550cf560a073748054f95da4741c994. The rebased candidate preserves manager 0.0.205 and executionenvironment 0.0.44, selects unclaimed CLI 0.0.122, and passes targeted real-RPC/consumer checks 2/2 plus the full local suite 270/270. No failed or retry-assisted rows/warnings occurred. CLI source pushed to final head cd057e8dc1afc3396fbbc6f10db0d7a62b356ce5; Both required Actions checks are green; build-watchman exited GREEN. Post-green test-comprehensiveness and adversarial-code passes are CLEAN on this exact head. Worker gates are complete; orchestrator owns merge. Direct consumer baseline pinned to protocol 0.0.568 fails 0/1 at the exact notice and is preserved at https://github.com/CodexCoder21Organization/kompile-cli/tree/wip/w1-cli-ownerless-baseline-20261008 @ cec81d3092d9a4803a25bf913f4475304a5418ea (intentional failing baseline, never merge). The old output-ordering question is settled by closeFuture listener completion before close-request promise completion, which provider.close awaits before output captures are restored.

Core main already registers the upload session ID as its stoppable run ID; the old missing-registration premise is refuted. The production client check prevents a following chunk after a response callback closes the workspace. The resumed WIP's commit description overstated two fixes: negative-index rejection and stopped-unfinalized terminal rejection were not implemented, and their unchanged tests failed. Implemented both. Its native-stop race tests then failed because sendDisposition caught an already-terminal outpace/supersede exception and sent a cancellation RPC; now the exact same-run, known-terminal, requested-operation rejection bypasses that downgrade and is accepted as completed cleanup. All four reproduce-before-fix scenarios pass 4/4. All 23 changed/neighboring targets pass across the 20-pass broad run and corrected cancellation trio; seven native-handler fixture constructors were updated honestly and exact first-close assertions preserved. Source candidate 2367f3ef03d47aeac340a6a59b817813c087715b selects manager 0.0.206 after an exact-coordinate 404. Core full local suite completed within its cap: 1218/1222 pass, four prepared-fixture failures because documented scenario warmup was omitted. Ran warmup into a new cache, set KOMPILE_SCENARIO_CACHE, reran all four unchanged failures: 4/4 pass. Source is unchanged; this is not a single 1222/1222 green run. Final source head2367f3ef03d47aeac340a6a59b817813c087715b is pushed. Required fat-jar is SUCCESS; required bld-all-tests is FAILURE, three shards passed and shard1 failed306/307 at testExplicitDirectoryKilledAttemptRecovery. The unchanged local targeted test passes6196ms; reliable reproduction and root cause are not established. Retry-assisted testExplicitBuildFailureKeepsFailedTestInResultsAndXml and testResultIndexPreservesObservableWorkspacePath also require diagnosis. Full shard XML/attempt histories/thread dumps and targeted result are on the evidence branch. No speculative fix or CI rerun. Post-green reviews remain pending. Local API finalization is synchronous and tests prove terminal publication ownership, not server worker interruption/draining.

Relevant PRs and recoverable state

| https://github.com/CodexCoder21Organization/kompile-cli/tree/wip/w1-cli-ownerless-baseline-20261008 | cec81d3092d9a4803a25bf913f4475304a5418ea | no PR; intentionally failing baseline, never merge | Direct CLI regression on old protocol 0.0.568 and full failure trace |

Apply recovery patches only after cloning fresh main at their recorded base (patch headers preserve exact source); all three final source heads are pushed; never accept old-head CI. Only ready CLI lock remains for the orchestrator; unready UrlProtocol/core locks are released at worker stop. Do not touch another worker's lock.

Published but not merged

Protocol 0.0.569 was published by the earlier session from https://github.com/CodexCoder21Organization/UrlProtocol/commit/64c672f9154f8880d3d92d47d66c5836ac3ba948, an earlier source head of the open fix. Its exact SHA-256 was freshly verified above. The current rebased source includes later main changes and is not claimed to be byte-identical to that old artifact. Native handler buildtest.server:buildtest-server-service:0.0.26709273972 downloaded SHA-256 80f098ccfefbcb863b3f58a78424ae6434f4781524e028df93217e7645cad277; no new publish or production change by w1. Historical CPU deployment is not reverified because all deploy work is excluded from this sweep.

Next steps

1. UrlProtocol full local is complete (2339/2339); finish its final required CI checks and reviews. Core required bld-all-tests is FAILURE on2367f3ef; do not requeue blindly. Shard1 testExplicitDirectoryKilledAttemptRecovery timed out at30000ms; unchanged targeted local passes6196ms. Full XML/thread dumps and attempt histories are durably saved in sweep-evidence on the linked evidence branch. Its attempts2/3 fail the initial no-owner-tree assertion because the first killed test child leaves a stable runtime tree; the initial timeout remains unexplained. Also diagnose retry-assisted testExplicitBuildFailureKeepsFailedTestInResultsAndXml and testResultIndexPreservesObservableWorkspacePath. A reliable local reproducer is required before any fix. The timed-out Coursier worker had existed only0.13seconds; its parked frame alone does not establish a stalled fetch. Core rules require bld-all-tests and bld-build-fat-jars; it has no kotlin.build (remote) check. The completed full-local run and warmup-corrected four failures are recorded above; require the final full four-shard aggregate. Never accept UrlProtocol remote red. Reproduce any test failure locally before fixing, not by requeuing CI.
2. Once each final head is green, do independent test-comprehensiveness and adversarial-code review passes on that exact head. Orchestrator alone decides merges after its own review.
3. After merges, publish manager 0.0.206 (or advance if newly claimed; never overwrite), using the Maven publication skill and SHA-256 byte verification. CLI's release/pin sequence must use current main, not old 0.0.117/0.0.118 numbers; publish CLI 0.0.122 as appropriate, then a new CLI manager-pin PR/release for the fixed manager.
4. Run five sequential real remote BuildTestWui apiEndpointReturnsJsonTest acceptance runs with the final released CLI, expecting 5/5 and no post-verdict thread-root notices. Update BuildTestWui scripts' CLI pins in a reviewed green PR. Acceptance and publication depend on merge decisions and cannot be completed by this worker before them.
5. Orchestrator completes the handoff when releases, pins, and acceptance are verified. Never deploy/restart/update routes/rebake as part of this sweep. No disk-exhaustion work remains in this handoff.

Operational notes

All heavy full suites use flock <sweep>/fullsuite.lock timeout --kill-after=15s 1800 scripts/test.bash --local --test . --log <log> and tracked outer deadlines. Remote submission is reported broken before tests with Unknown upload session, so use local full suite and required CI as specified in the worker brief. The long baseline checkout path exceeds the default cache basename limit; invoke its generated launcher directly with an explicit short --cache-location. Old gh pr edit fails on Projects classic GraphQL, so PATCH the REST pull-request endpoint with the exact JSON body file. These challenges are recorded in findings for the orchestrator; report-challenge CLI auto-merges and is prohibited in this worker. All artifact caches/JARs stay outside source commits.

OBSERVED (13:13 UTC): Fabric host logs confirm remote run 94b3d806 is executing; shard-zero droplet is provisioned and awaiting guest readiness. Host stat reports build.log and test-events.jsonl modified at 13:13:27 UTC (6909 and 458301 bytes). WUI HTTP 503 while rebuilding a cached results view is a monitoring limitation, not an executor-failure verdict. No production mutation.

OBSERVED (13:26 UTC): Filtered durable run.json returned BUILDING, testsTotal=2339, testsPassed=0, testsFailed=0, refuting cached WUI PENDING. Fabric refused the entire 2.5MB metadata file under its daemon read budget, so a read-only Python field projection returned a small response. No daemon heap setting or production state changed. Full-suite failure traces and corrective results are saved at evidence branch 23a298233acd1c89a70a97eec748f4d3e79b16e9.

Release ordering is an actual dependency from the original handoff: publish CLI “from the merge commit” and manager “from main”, then make the new manager-pin CLI PR. The worker cannot satisfy those prerequisites by merging or publishing an unreviewed candidate. BuildTestWui fresh origin/main still pins CLI 0.0.114 in scripts/build.bash and scripts/test.bash, so its final pin step is still required.

Worker stopping state

PARTIAL at the sweep timebox. CLI is the sole ready PR. Core source is durable and unchanged after CI failure; full failure traces and unchanged passing targeted result are saved remotely. No independent post-green review is claimed for core or UrlProtocol. Continue core diagnosis with a deterministic kill/re-entry experiment against the public scenario helper, separately measure the preceding operations that exhausted the first attempt budget, and reproduce the two retry-assisted cases locally. Do not raise timeouts, reduce iterations, weaken assertions, or accept a red required check. Continue UrlProtocol final remote watch with build-watchman; perform both reviews after green.