id: hf-2026-09-27-finish-tokened-ownership-regressions-and-get-buildtestembedded-s-four-shards-green url: url://handoff/handoffs/hf-2026-09-27-finish-tokened-ownership-regressions-and-get-buildtestembedded-s-four-shards-green title: Finish CI for https://github.com/CodexCoder21Organization/BuildTestEmbedded/pull/1204; then final-review and merge summary: Finish the existing CI run, then supervisor final review and merge decision. F2 PR is open at 33c4558001f05f93eec32280e83d5afabc150e43; all local gates complete (original 5/5, family 23/23, current fail-first red and restored green), lane 27 passed / 1 expected old-source failure. Four shards remain IN_PROGRESS; no CI-green claim. created: 2026-09-27T15:29:56.057Z completed: null dependencies:

Historical snapshots (superseded; kept for provenance)

Quota-stop snapshot 2026-09-28 18:13 UTC

RE-VERIFY banner (added 2026-09-28 18:25 UTC by fable-code-2026-09-28): everything below this banner is a write-time snapshot taken when the supervising session stopped because its codex delegates hit the account usage limit (reset 2026-10-05 05:43 UTC); re-check every PR with gh pr view <n> --repo <owner>/<repo> --json state,mergeStateStatus,headRefOid,statusCheckRollup and every branch with git ls-remote before acting. Session artifacts (lane briefs lanes/*.md, findings out/*-findings.md, review verdicts out/r*-verdict.md, the supervisor's review log out/supervisor-reviews.md, slot queue out/followups.tsv) are at https://github.com/CodexCoder21Organization/PlanRepository/tree/handoffs-artifacts/2026-09-28-fable-board-drain-quota-stop/handoffs/artifacts/2026-09-28-fable-board-drain . The board-level resume procedure lives in the handoff titled "Resume the open-handoff drain". The shared droplet service dropped its RPC connection twice (17:13?17:18 and again about 18:04 UTC), failing every remote run in those windows before any test executed; runs that report zero tests executed with a provisioning error are infrastructure verdicts, not code verdicts.

State: https://github.com/CodexCoder21Organization/BuildTestEmbedded/pull/1073 OPEN/CLEAN at e495ff4f (branch fix/droplet-deletion-requires-id-ownership), all five Actions checks green. The supervisor's final review (artifacts out/supervisor-reviews.md, 17:24 UTC) returned two findings that must be fixed before the delegate reviews run: F1 the PR reversed the documented safety rule that an unreadable run record never authorises deleting a droplet (BuildDropletReconciler now reaps on Unreadable when no provider timer and past grace; the service comment at BuildTestEmbeddedService.dropletRunLookup still says the opposite; the test reapsDropletWhoseAutoTerminationTimerNeverRegistered was rewritten to expect the delete) ? restore keep-on-Unreadable; F2 recoverFromFailedCreation now fails a successful adoption when deleting a duplicate throws ? keep the adoption and surface the failure without failing the allocation (the quota-path throw stays). Fix lane o06g (brief lanes/o06g.phase.md) got as far as wip/o06g-quota-stop-20260928 at 2284ff5c: the restored test expectation and a new fail-first duplicate-adoption test (first run cce56fcd failed to compile because the test accessed the internal ownsBuildDroplet; the corrected run 18ac37ab died in the 18:04 provisioning outage). No product change yet. Next steps: get the fail-first verdicts for both tests on the unchanged head, implement F1 and F2, run the reconciler and DropletManager selectors remotely, rebase, push, refresh the PR body, keep the five checks green, then both reviews (lanes/r1073.md is drafted), supervisor final review, enqueue.

B73 timebox checkpoint, 2026-09-27 22:35 UTC

RE-VERIFY: This is a write-time snapshot. B73 stopped at an incomplete source checkpoint; the original PR is not ready. The sweep4-B73 claim will be released with a BLOCKED report. No merge, enqueue, rerequest, deploy, publish, or remote tests were done.

Remaining work first: diagnose and repair the remaining branch regressions listed in the B73 full-suite analysis, prove each on its public test 5/5 first attempts, rerun the 129 original selectors together, then acquire the machine-wide suite slot for one new exact-head full local suite. Run both final-head reviews. Only when local gates are green may the original PR be updated under the repo lock and its bld-build plus four shards watched. The A13 unreadable-owned-run keep-after-grace rule stands. Do not shorten the 64-character test ID; the recommendation is in needs-decision/B73-name-budget.md.

Branch	Verified remote SHA	State
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/b73-1073-ownership-2026-09-27	89a0572569c5fc8b74f56304c848559af016eeaa	Clean source checkpoint on main 50c792c8b. Contains worker generation reset, late-quota exact-name ownership, ledger failure cleanup retention, and tokened fixture updates, including two unreadable-record retention fixtures that passed 5/5 first-attempt local runs. Known targeted failures remain.
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/b73-1073-evidence	2ad28a2304be59e31eedfaa3d1f5e896b643bc1f	`reports/B73/`: full-suite XML, main comparison XML, targeted logs, analysis, findings, reviews, resume instructions.
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/1073-W92	c2c10b5fcf2b7c1600aa4c89c44616157c224792	Prior W92 source checkpoint.
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/1073-W92-evidence	2ee3da665645118083df2a08e785fd726f609f9e	Read `reports/W92/report.md` first for prior diagnosis.

The eight W92 broad-gate failures passed first actual local execution after the rebase compile repair. The combined original-selector gate passed 129/129 with zero retries at b26172c2. B73's one full local suite on exact pushed head 19e12d4bd was 2842/2860 pass, 18 fail, plus seven retry-passes; the shared suite slot was released at 22:09 UTC. All 17 matching failing selectors passed first attempt on current main 50c792c8b. Subsequent targeted fixes are captured on the source branch; there has been no second full suite; only the two unreadable retention fixtures have final-head 5/5 proof.

https://github.com/CodexCoder21Organization/BuildTestEmbedded/pull/1073 was reverified OPEN at old head db0c1ab6e14c768f6c2b6f5f817ad3defcc04abf at 22:27 UTC. B73 did not push to it. See B73 resume details and full traces before editing. The local branch and evidence branch were pushed with --force-with-lease under the repo lock and verified by git ls-remote.

B73 checkpoint, 2026-09-27 21:31 UTC

RE-VERIFY: This is a write-time snapshot. Recheck the PR, remote branches, suite lock, and CI before acting. The active claim is sweep4-B73.

Remaining work: finish the running full local suite from the exact pushed source head, compare any failures by name to main, then update the original PR and watch bld-build plus four shards. B73 acquired the machine-wide suite slot at 21:30:47 UTC on attempt 32; the EXIT trap releases it. Never merge, enqueue, rerequest, deploy, publish, or run remote tests.

Branch	Verified remote SHA	Contents / state
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/b73-1073-ownership-2026-09-27	19e12d4bd1db9b72d70cc340cf669d9787b94dbe	W92 rebased on main a0f0fde76, with Okio port, quota-state repair, confirmed-deletion repair, and fixture corrections. Local 129-selector gate passed; full suite running on this exact pushed head.
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/b73-1073-evidence	535909b59f5c1af4a5340c2e9b6e9837e16af63f	`reports/B73/`: fail-first and passing XML, gate selector list, findings, reviews, and provider-name recommendation.
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/1073-W92	c2c10b5fcf2b7c1600aa4c89c44616157c224792	Prior source checkpoint, now continued on the B73 source branch.
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/1073-W92-evidence	2ee3da665645118083df2a08e785fd726f609f9e	Prior W92 report and broad-gate traces.

https://github.com/CodexCoder21Organization/BuildTestEmbedded/pull/1073 was OPEN at old head db0c1ab6e14c768f6c2b6f5f817ad3defcc04abf when rechecked at 21:09 UTC; the previous bld-build and four shards remain failed. B73 has not updated it. Nothing was deployed or published by this lane.

The eight failures handed over by F11/W86 each passed their first actual local run after the rebase compilation repair. The W92 original 128 selectors and one new public deletion-worker test passed together 129/129 at b26172c2, with zero retry entries; 19e12d4bd is only a whitespace correction. The initially failing 128 gate had three additional failures. Startup setup had persisted TESTING runs with zero provider IDs despite creating locally owned tokened droplets; after assigning the real IDs, it exposed a duplicate provider-delete race (34 successful calls for 32 IDs). A public worker test deterministically failed before the fix and passed after: confirmed IDs are now recorded under the same intent lock used by enqueue. The shard fixture now checks replacement before final cleanup, and the quota fixture distinguishes the primary tokened name from its sibling. A conclusive quota refusal clears its in-flight flag only after a complete cleanup observation; ambiguous create responses remain unresolved. Main's full unreadable-ledger peer-survival fixture is restored. A13's unreadable-owned-run rule remains unchanged. The 64-character name case passed unchanged; reports/B73/name-budget.md recommends a bounded provider base with the full ID in the ledger.

Both independent final-head reviews, including the tokened-record resource table, are in reports/B73/review.md. The suite waiter fetched/rebased main just before the build and found no head change; git ls-remote matched the pushed WIP head, and the working tree was clean. If the suite passes, reverify PR OPEN and unqueued, fetch/rebase under the repo lock, push with --force-with-lease to fix/droplet-deletion-requires-id-ownership, rewrite the body why-first with the CI regression link and current gate counts, then watch the five checks without merge-queue entry. Capture any full-suite result and final PR state in this handoff before stopping.

W92 checkpoint, 2026-09-27 18:07 UTC

Remaining work: diagnose eight broad-gate failures, validate the full ownership gate on the fixed source, update the original PR and get four shards plus build green. This is CHECKPOINT, not ready for merge. Claim codex-W92 remains held. No merge/enqueue/deploy/publish.

Branch	Remote SHA	Contents
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/1073-W92	c2c10b5fcf2b7c1600aa4c89c44616157c224792	W86 rebased on main1b80be83c, plus complete-observation ownership resolution and two public regression tests.
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/1073-W92-evidence	2ee3da665645118083df2a08e785fd726f609f9e	reports/W92/report.md, findings, separate reviews, red/green logs/XML, all broad results and full failure traces, selector inventories/scripts.

Read reports/W92/report.md FIRST. New mechanism measured: a reclaimed lost-response allocation retained name-only authority, so a later peer with the same tokened name was incorrectly deleted after25h/restart. Both public tests failed on old source, then passed2/2 at finalc2c10b5fc. Eight-case neighbor batch passed8/8 on the same production source, nine distinct passing tests total. Successful complete observations now narrow inactive unresolved records to all observed IDs; empty/failed lists and live attempts stay unresolved. Separate final test and source reviews found no further concrete source issue, but broad runtime gates remain unfinished.

Prior128 gate: https://buildtest.kotlin.build/run?id=620e9eee passed128/128 at e0740571 BEFORE the new fix. Broad first280: https://buildtest.kotlin.build/run?id=b59f1bda finished272PASSED/8FAILED at that same pre-fix head. Complete final pagination has zero missing selectors/errors. See b59f1bda-failures-latest.md for all full traces. Seven failures are preparation/follower cases; the eighth is ambiguousWriteFailureCleanupListingThroughSandboxBridgeReentersAdmission. No new cause has been established for those eight failures. Do not infer that concurrent q07 work fixes all of them.

Next steps:

Reverify refs/main and inspect q07 preparation work at https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/q07-preparation-2026-09-27 (observedc5f705472ef34102f84b896a34b4ac7f277dc08d). Do not overwrite it. Diagnose the eight failures using full traces and deterministic tests. Known unrelated fixes owned by other lanes should be coordinated, not duplicated.
Rerun128 after the new fix and finish gate-all-current.txt (943selectors). Existing disjoint broad files contain280/280/253 beyond128; two new regressions are separately in the total inventory. Broad2/3 were never submitted. One remote client and one slot-wrapped local JVM, never local full suite.
Both review passes again after any change, final tests, verify original PR OPEN/unqueued, rebase main, force-with-lease update original branch and why-first description, watch four shards+bld-build with build-watchman and STOP at green. Never merge/enqueue.

At18:03 original https://github.com/CodexCoder21Organization/BuildTestEmbedded/pull/1073 stayedOPEN,DIRTY,unqueued atdb0c1ab6e14c768f6c2b6f5f817ad3defcc04abf; allfivechecksFAILURE. It was not updated. Coordinate0.0.69274740 remained404. Source working tree clean. Read-only watcher may still finish observing the now-terminal run; all test clients have exited. External client interruptions and intermittent503 result pagination are recorded, not treated as test causes.

Historical checkpoints (superseded where W92 gives newer evidence)

W86 continuation, 2026-09-27 16:43 UTC

RE-VERIFY: W86 is stopping at a CHECKPOINT after completing durable engineering work. codex-W86 retains the claim for the orchestrator. Do not merge or enqueue; only the orchestrator may do so. PR and refs were checked at 16:41 UTC.

Remaining work: diagnose the two intermittent prior-gate failures, run the complete ownership-related gate, then update the original PR and get all four shards plus bld-build green. This is not READY.

Branch	Remote head	Contents/state
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/1073-W86	24efce2ca0b567a84a5223d16b0f4293ab4d0eac	Source rebased on main 8d76d002; conditional timer reaper, bounded long names, five fixture migrations, recovery-list count correction, heartbeat diagnostic correction, stronger review cases, new post-create callback fix. Final stable head passes 8/8 targeted cases.
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/1073-W86-evidence	9cb63749c8b80ba6615d6b44be3ab907e31c1743	reports/W86: paginated verdicts, full traces, fail-first/green XML, reviews, findings. Final report, fail-first callback trace and passing 8-case XML are durable here.
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/1073-F11	acfcb76866ec3e2e18a42d954bc15dbf25a587a3	Previous source checkpoint; W86 descends from it.
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/1073-F11-evidence	38fcfa6dd766e7a3d4ceb64bb0b15da6eba09628	Previous evidence and broad selector files.

PR https://github.com/CodexCoder21Organization/BuildTestEmbedded/pull/1073 remains OPEN, unqueued, DIRTY on db0c1ab6e14c768f6c2b6f5f817ad3defcc04abf; all four bld-test-shard checks and bld-build FAILURE. The PR branch is unchanged. main is 8d76d00240dc48a84697c9545c4ac5ca1d42af2d. Coordinate 0.0.69274740 was checked unpublished; nothing published or deployed by W86.

The orchestrator's unreadable-run decision is implemented: keep when positive provider auto-termination deadline is registered; with no timer, reap an owned droplet after known-age grace. Unknown age and all peers stay. Four DropletInfo conversions now preserve the deadline. Expanded scheduled-cycle test measured red 0/1 then green 1/1; remote gate also passes.

Long-ID mechanism: new 56-character rejection broke formerly accepted run IDs. Bounded hashing includes the full run ID and shard suffix while the durable ledger keeps original identity. Prefix cleanup now uses that record. New long-ID/restart/shard test measured red 0/1 then passed remotely.

Regression run https://buildtest.kotlin.build/run?id=c530dfe1 completed 125 passed / 3 failed of128selected at candidate 9764846a5. All five corrected fixtures and F11's dynamic-survivor fixture pass. Three failures: heartbeat diagnostic expected seconds but actual 15 minutes (fixed and local1/1pass); resumedShardedRunStillFailsWhenNoShardCanProgress (remote 30s timeout, local 7.4s pass); testRecoveredMalformedFingerprintSpillPreservesEarliestSequence (remote 10s attachment wait, local 9.8s whole-test pass). The last two remain unexplained; do not infer a fix from one local pass. Remote CLI lost a getBuildRun response and automatically attempted cancellation, but read-only API confirms all 128finished. Do not repeat cancellation.

Inherited broad-3 run https://buildtest.kotlin.build/run?id=6ac01163 completed258 passed / 12 failed. Full failures are in reports/W86/broad3-failures.md. One fixture's recovery-list count included the removed pre-create observation and is corrected but targeted verification passed. Other failures remain undiagnosed, including shared-preparation follower waits, attempt-history/archival timeouts, and mixed-release cache tests. Shared-preparation q07 has active concurrent work; inspect that branch before touching the same code.

Review found a further deterministic bug: a one-shot caller callback exception after provider creation was caught as a provider failure, recovered by name, and returned success. New allocationCallbackFailureAfterCreatePreservesOriginalException fails with complete provider trace in review-gaps.log. d42c22f1c records the callback's original exception and propagates it before provider classification/recovery, retaining interruption restoration. The exact final head passed an eight-test batch under jvm-slot.sh, including callback before/after create, long service deletion, name-length boundary, admission recovery, interruption, quota recovery and quota exhaustion. Mixed opaque-ID restart tests passed 2/2. The new long-service fixture initially hit the live-lease guard; it now seeds and advances the same ManualClock and passes without changing the contract.

Next steps in order:

Read reports/W86/report.md and both review reports on the evidence branch, then reverify refs/PR. Both review passes have no remaining concrete finding; all local test invocations and remote watchers have finished.
Reproduce the two remote intermittent failures locally with honest concurrency/clock ordering and capture complete state; no timeout increases or speculative mitigations.
Diagnose remaining broad failures. Preserve concurrent actors' work. Run all required selectors: gate-w86-all.txt is the conservative940-selector union of the current grep surface, the original937-ish gate, and added tests. Split into manageable remote batches; one at a time. Keep failing cases as failures, not filters or skips.
Repeat both review passes on the final head. Ensure all prior failures pass in one remote batch and every broad gate passes.
Verify original PR OPEN and unqueued, fetch/rebase main, run final tests, push --force-with-lease to its branch, refresh why-first description, and watch four shards+bld-build using build-watchman without --to-merged. Stop at green.

Tool notes: export PATH=$HOME/.local/bin:$PATH; local JVM commands use cx/jvm-slot.sh, one local invocation at a time. Never run the full suite locally. build-watchman0.0.18 only counts its first100 result rows, including FILTERED_OUT; collect /api/test-results?id=...&limit=500&page=N and count status. Handoff RPC can report Stream closed; retry up to3. Evidence push used git -c http.version=HTTP/1.1 after an HTTP/2 stream error. No production actions.

Historical F11 checkpoint

Handoff: Finish the tokened ownership regression fixes and get BuildTestEmbedded's four shards green

Written 2026-09-27, F11 checkpoint after the requested approximately 80-minute work period.

RE-VERIFY: This is a snapshot. Check the PR with gh pr view 1073 --repo CodexCoder21Organization/BuildTestEmbedded --json state,headRefOid,headRefName,statusCheckRollup and query GraphQL isInMergeQueue. Check remote test runs at https://buildtest.kotlin.build/api/runs?limit=100 and all pages of /api/test-results?id=<run>&limit=500&page=N. FILTERED_OUT means not executed.

Mission

The user asked to fix regressions introduced by https://github.com/CodexCoder21Organization/BuildTestEmbedded/pull/1073 at head db0c1ab6, at cause, and get four test shards plus bld-build green. Do not enqueue, merge, publish, or deploy. Do not increase timeouts, reduce iterations, skip or weaken tests. Remote batches only; a single local diagnostic may run under cx/jvm-slot.sh. Stop at green. Before updating the PR, every original failure must pass in one remote batch, and every test whose file matches git grep -l -E 'buildtest-[a-z0-9]|ownership|Reconciler' tests/ must pass in remote batches. Both review passes must cover the final head.

Found and done

Read README, W78/W74 history and testing/philosophy guidance. Fresh clone at workspace/F11-bte, rebased onto main. All four logs from https://github.com/CodexCoder21Organization/BuildTestEmbedded/actions/runs/36322820773 yielded 123 distinct failures: shards 1/2/3/4 had 30/20/44/29. Complete logs and failures.tsv are preserved in the evidence branch.
The parser rejected hyphenated run IDs and misinterpreted run IDs ending in -sN. The four reported disappearing-record failures were not a concurrent deletion: validation failed before a record existed, then retry publication tried to update the base name. Owned name recovery now reads the exact durable runId; legacy parsing remains discovery only. Shard fixtures now pass the base runId with nameSuffix separately.
A distinct real retention defect existed: after restart, an empty fleet listing and 24-hour age removed an unresolved attempt record. New public test failed before the fix and passed after it. Prune now retains unresolved/unreadable records; ordinary resolved records still age out.
Added opaque run-ID restart/reconciliation tests, including empty, blank, slash, newline, and percent characters. Records for ordinary names keep their existing filenames; other names use bounded SHA-256 file keys while JSON retains exact identity. Prune verifies the record name maps to that file. Present blank IDs round-trip; malformed non-string record IDs are not reparsed. These tests passed remotely after failing before the fixes.
Corrected missing observation-error detail to fully qualified exception class and explicit missing-message text. Migrated name assertions, provider boundary classifiers, exact diagnostics, list-call counts, and ownership fixtures throughout the suite. Rationale is per-file in test-migration.md, rpc-test-migration.md and remaining-name-audit.md. No timing/iteration changes. A final test review found a same-name peer fixture that had stopped exercising an exact name collision; fixed at 25ff5ec44.
The user premise about unreadable runs is contradicted by the existing tests and README: they explicitly require keeping an unreadable owned run even after grace. An async question asked whether to change that contract to reap after grace. No answer arrived. The task explicitly requires approval before changing this test; policy remains unchanged. This decision is still required.
A callback-exception hypothesis was refuted: new callback identity test passed unchanged production. No production callback change was made.

Verification and remaining failures

Remote old-code run e0958a5e: 2 FAILED, 2794 FILTERED_OUT. Both new ownership regressions failed for the intended mechanism. Local opaque-ID and retention tests each changed from 0/1 to 1/1 after their fixes.

Intermediate remote run 0ac8e8b5: 58 PASSED, 5 FAILED. This snapshot predated fixture migrations. Four failures correspond to fixtures subsequently changed. testRecoveredTenShardWorkStateSharesProjection timed out at its first-idle latch; it later passed in the complete prior-failure batch, so its mechanism remains unknown, not proven fixed. A later local diagnostic was stopped at checkpoint without a verdict to release its host slot.

Required prior-failure batch fec57557, source head 48448b85c: 119 PASSED, 8 FAILED, 2671 FILTERED_OUT. This includes all 123 original failures plus four new tests. All three new ownership tests and callback coverage passed. Full errors are in reports/remaining-failures.md on the evidence branch:

dropletRpcCallerInterruptionIsPropagated
e2eTerminalReservationReleaseIsIdempotentAfterLateWorkerExit
e2eTransientWebcronTimerFailureRetriesDropletCreation
residentRunnerEventsDoNotReadRunJson
resumedShardedRunNeedsDynamicSurvivorBeforeReplacingMissingShards ? exact message fixture corrected later at acfcb7686, not yet tested. Root review added a named regex backreference so both message positions must use the same captured attempt name.
resumedShardedRunStillFailsWhenNoShardCanProgress
startupReadyFollowerDispatchesDuringHeldPlanJournalRead
testRecoveredMalformedFingerprintSpillPreservesEarliestSequence ? base name is 58 characters, but the attempt token reserves 8 of the provider's 64. Needs an explicit contract decision/fix at cause; do not simply shorten the test ID.

The other seven remain undiagnosed. Full stack traces are preserved; do not infer one shared mechanism without evidence.

Broad gate: 864 selected tests including new ownership cases, union with original failures gives 937 selectors. Four batches: 127 prior/new plus three disjoint 270-test batches. Initial broad submissions 743e3812/2a9fe325/2d499792 ran no tests because four function selectors used buildtest.embedded instead of community.kotlin.buildtest.projection.tests. Selector files are corrected from discovered projectName. Corrected broad submissions: da4182ff failed uploadChunk after five automatic client attempts; 6ac01163 was BUILDING and a405b0dc CANCELED at checkpoint. The broad2/broad3 client logs lack IDs, so establish their mapping by comparing selected test sets. Both clients were externally killed; do not assume the surviving remote run stopped. Recheck before submitting anything else.

Local character-ID green invocation failed in the 1200-second build-script compilation budget, before executing the test; the same test passed remotely. This is not evidence against the ownership fix.

Independent production review found and then verified the blank-ID/hash-path fix, clean at 48448b85c. Test review verified the collision fix at 25ff5ec44. The later aggregate-message fixture at acfcb7686 still needs a final review and test. No final-green claim is made.

Relevant refs and unmerged state

Ref	Verified state
https://github.com/CodexCoder21Organization/BuildTestEmbedded/pull/1073	OPEN, isInMergeQueue=false, still db0c1ab6e14c768f6c2b6f5f817ad3defcc04abf. Four old shards and bld-build failed. PR branch untouched.
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/1073-F11	Code checkpoint acfcb76866ec3e2e18a42d954bc15dbf25a587a3, pushed and clean. No stash or extra worktree.
https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/1073-F11-evidence	Verified remote head 38fcfa6dd766e7a3d4ceb64bb0b15da6eba09628. Separate evidence-only branch: reports, selector files, and evidence.tar.gz with all full logs/API snapshots. Do not merge this branch into source.

Nothing was published or deployed in this effort. All actions were source/test changes, remote test submissions, read-only state queries, WIP pushes and checkpoint writes. Coordinate 0.0.69274740 was retained; its POM returned 404 at the initial check. Recheck before using it later.

Next steps

Re-verify the PR, both remote branches, and all remote run states first. Recover the source and evidence branches; unpack evidence.tar.gz for exact selector lists and full logs.
Obtain the user's unreadable-run policy decision. Preserve the existing keep behavior until answered.
Diagnose each remaining regression through a public failing test. Preserve all timing budgets, counts and full-message assertions. Recheck the aggregate-message fix and same-name peer fixture on the final head. The unrelated shared flake projectionStartupRetentionGapKeepsInWindowReadersLive belongs to lane F9; report it, do not diagnose it.
Finish the broad gate, including any remote run still executing. Corrected selectors are gate-broad-{1,2,3}.txt; run-gate.py builds repeated --test arguments and writes XML/logs. Every prior failure must pass together in one final remote batch, not by combining separate passing attempts.
Repeat both final-head reviews. Fetch/rebase main before builds. Verify PR OPEN and unqueued immediately before pushing with explicit force-with-lease to fix/droplet-deletion-requires-id-ownership. Keep the reserved coordinate unless already published. Add a description line linking the regression run and explaining the fixes.
Watch with coursier launch buildwatchman:build-watchman:0.0.18 -r https://kotlin.directory -- --repo CodexCoder21Organization/BuildTestEmbedded --pr 1073 --interval-seconds 300. Never --to-merged. Stop at all four shards and bld-build green; no enqueue, merge, publish or deploy.

Operational notes

Local root: /tmp/claude-1000/-code/1f0f7d17-dc24-4d3a-a8ff-48753bfab06a/scratchpad. Code at workspace/F11-bte; evidence at cx/lanes/F11. rg is unavailable; use git grep/Python.
Single local test only: ../../cx/jvm-slot.sh scripts/test.bash --local --test NAME from the checkout. Two host-wide slots. Never a local full suite.
Remote API pages must exclude FILTERED_OUT. Build-watchman 0.0.18 incorrectly counted its first 100 FILTERED_OUT rows as completed and printed misleading PENDING/100-of-100 warnings; API pagination is the workaround.
Remote clients can drop after submission. Read the run's API state instead of resubmitting blindly.
Ecosystem friction is recorded in findings: watchman filtered-result counting, uploadChunk timeout, and local compilation budget failure. report-challenge CLI was not invoked because it automatically enqueues/merges, which this task forbids.

bte-f2b checkpoint (2026-10-02 16:21:41 UTC)

RE-VERIFY: This new checkpoint supersedes the unexecuted-test state above. Re-check all PRs and remote branch heads before acting. No merge, queue, deployment, publication, or handoff completion was performed.

Main e04680a0429b5dad204adaa966fff19f7eedcd33 still reproduces F2. The original unchanged test executed and failed in 2.1 seconds with Failed to destroy duplicate droplet 101 ('buildtest-adopt01-a4cbf62'): java.io.IOException: provider rejected duplicate delete, at deleteDuplicateDroplet:2070 and recoverCreatedDropletAfterFailure:1970. The console showed adoption 102 selected before deletion. Baseline runner: 0/5 pass, 5 failed; four failures were draft compile errors (Unresolved reference: close), not product verdicts.
Corrected the four draft FakeFileSystem.close() calls only, retaining checkNoOpenFiles and all behavioral assertions. Original regression is unchanged. Full baseline stacks and test-design review are in reports/bte-f2b on the checkpoint branch.
A corrected old-source rerun waited for both compile slots and was stopped before test.bash began so the now-proven owner fix could be verified next. No second baseline verdict exists. The owner change is locally committed with its first five-case verification queued; do not infer a passing fix from this checkpoint, which contains only tests/evidence.
Remaining delegate work: execute the fix's five-case run, prove original F2 5/5 plus the 23 family selectors, final reviews, push the fixed head, open a PR and watch one CI run. The slot helper is now /tmp/claude-501/-code/4719c767-3557-4cb1-90f8-28975c4c3602/scratchpad/withlock.sh, with two slots and a 3000-second bounded wait. Do not use the obsolete single-lock command in historical next steps.

Repo	Branch	Verified remote SHA	PR	State
BuildTestEmbedded	https://github.com/CodexCoder21Organization/BuildTestEmbedded/tree/wip/bte-f2b-2026-10-02	66e0e489fbd7d562b300a981d41175a0a1d9fe44	none	Confirmed F2 fail-first evidence, original regression and four corrected guard drafts; no production fix on remote yet

Note 2026-10-02 21:45 UTC from fable-drain-4719c767 (user stop request)

A lane on this handoff was stopped mid-work by a user stop request; it held no claim. Its findings are under https://github.com/CodexCoder21Organization/PlanRepository/tree/wip/fable-drain-4719c767-artifacts-2026-10-02/handoffs/artifacts/fable-drain-4719c767-2026-10-02/findings (kee72: kompile-executionenvironment PR 72 scoped verification complete, remote verdict was being watched; rev-bte1041: delegate review of https://github.com/CodexCoder21Organization/BuildTestEmbedded/pull/1041 at d2f274f8 had just started). The campaign's resume plan is in handoff hf-2026-09-28-resume-the-open-handoff-drain-land-the-eight-in-flight-prs-through-reviews-and-merge-then-continue-the-40-item-queue.