Repository navigation
Conversation
Contributor
|
🦞👀 Pull request received. I will update this pull request when review starts. ClawSweeper review completeClawSweeper finished reviewing this revision. The review result is being finalized. |
openclaw-barnacle
Bot
added
docs
clawsweeper
Bot
added
P2
Contributor
|
Codex review: blocked before merge. Reviewed October 7, 2026, 4:20 PM ET / 20:20 UTC (Revision 46). ClawSweeper reviewWhat this changesThe branch expands opt-in AWS runner placement, splits measured CI test tails, moves admitted control jobs to Blacksmith, and repairs subprocess diagnostics, Gateway scheduling, and test fixtures. Merge readiness⛔ Blocked before merge - 5 items remain The central routing expansion remains absent from main, and this protected, maintainer-owned branch retains distinct, measured work. It is not a cleanup-close candidate. Priority: P2 Review scores
Verification
How this fits togetherOpenClaw's CI planner combines selected tests, workload measurements, backend settings, and admission checks to choose runner jobs. Those jobs execute validation and feed the aggregate merge gate. flowchart TD
A[Selected tests and measurements] --> C[CI planner]
B[Backend and admission checks] --> C
C --> D[Runner placement and packing]
D --> E[Hosted and Blacksmith jobs]
D --> F[Opt-in AWS jobs]
E --> G[Validation results and merge gate]
F --> G
Before merge
Findings
Agent review detailsSecurityNone. PR surfaceSource +62, Tests +3793, Docs +345, Config +75, Other +389. Total +4664 across 41 files. View PR surface stats
Review metrics
Stored data modelPersistent data-model change detected: Merge-risk optionsMaintainer options:
Technical reviewBest possible solution: Keep hybrid as the operating default and qualify the expanded route through the existing planner, preserving complete coverage, resource limits, fallback behavior, and newer owner repairs. Do we have a high-confidence way to reproduce the issue? Yes for the remaining documentation defect: the pinned branch removes a heading fragment that the runners page still references. The broader routing goals were assessed from recorded native runs, without executing target code. Is this the best way to solve the issue? The existing planner and lifecycle owners are appropriate boundaries, but acceptance remains incomplete until integration is refreshed and whole-run qualification establishes the intended result. Full review comments:
Overall correctness: patch is correct AGENTS.md: found and applied where relevant. Codex review notes: model internal, reasoning medium; reviewed against e73028f81cda. LabelsLabel changes: No label changes. Label justifications:
EvidenceWhat I checked:
Likely related people:
Rank-up movesOptional improvements that raise the rating; they are not merge blockers.
Rating scale
Overall follows the weaker of proof and patch quality. Workflow
HistoryReview history (45 earlier review cycles; latest 8 shown)
|
clawsweeper
Bot
added
proof: sufficient
steipete
force-pushed
the
ci-burden/s2-runson
branch
from
September 23, 2026 06:43
c2bca82 to
1b46053
Compare
clawsweeper
Bot
added
status: 👀 ready for maintainer look
Contributor
Author
|
TP-G1 coordination from #156270: the real-Gateway inventory now has two planner-owned rows (6/17 files, 7/32 cases, desktop proof once), with unchanged worker limits and runner labels. Exact-head main qualification 35851552173 at The split is balanced: two standalone companions run together after the desktop row's serial phase, without a bundled preview build. The build still costs 87/81 seconds (bootstrap validation 37.1/31.1); UI steps 216/350 seconds, with desktop 122 seconds only in row 1. SDK declarations are already removed from these runtime-only consumers and remain validated in build-artifacts. No timeout, retry, coverage, runner-label, variable or ruleset changes were used to obtain these results. Your current routing head deliberately retains real-Gateway on Blacksmith. Please qualify a route with higher clock speeds for these two rows through S2's runner owner, using the S2b m8azn.3xlarge evidence or the cheapest measured class that fits. TP-G1 will keep the split in planner data and leave label/backend policy with S2. The two-worker bound and exact file inventory must remain unchanged, and the final native jobs must each pass within 450 seconds. No auto-merge is armed on TP-G1. |
steipete
force-pushed
the
ci-burden/s2-runson
branch
from
September 23, 2026 11:27
1b46053 to
791e06f
Compare
clawsweeper
Bot
added
proof: sufficient
steipete
force-pushed
the
ci-burden/s2-runson
branch
from
September 23, 2026 14:37
791e06f to
0e51d65
Compare
Replace the single AMD capacity pool with prioritized instance alternatives across the configured RunsOn subnets. Admit Spot only when a known forecast plus measured startup allowance fits eight minutes; keep interruption retries disabled and use on-demand otherwise. Split the underpriced CLI runtime cohort without dropping coverage or changing workers, deadlines, caps, or hosted stripe routing. Retain the unresolved GitHub-projects UI spec in a disjoint Blacksmith row and pin AWS Node for comparable runtime evidence.
Creation responses marked sessions active without registering their run IDs, so terminal events could leave stale running state. Materialize identity and run receipts through the shared fixture owner, preserve early terminal or abort receipts, and discard unrelated state when the session generation changes. Extend serialized RPC coverage for creation, replay, replacement, and early settlement. Include the awaited RPC method in existing timeout errors without logging request payloads or changing deadlines. This repairs proven fixture defects; the historical AWS timeout remains isolated rather than declared fixed.
The RunsOn section rewrite accidentally removed the following hosted-assignment guidance and Related links. Restore that existing tail byte-for-byte so the assignment-guard anchor resolves and the separately owned hosted routing documentation remains intact. Validation: markdownlint scanned 1306 files with zero issues; docs listing, formatting, diff checks, and P2 review pass.
Direct AWS proof used a different operating-system image and could not qualify the unexplained native RunsOn RPC timeout. Add a main-qualification-only counterpart with the retained spec's exact selector and worker policy, and record both providers' actual toolchains. Production routing and retry policy remain unchanged.
The measured 85-file tooling envelope took 786 seconds on Blacksmith's two visible CPUs and retained that placement in the RunsOn PR plan. Move only long serial two-worker ordinary tooling to the existing diversified general pool, preserving all selectors, packing, workers and deadlines. Its forecast keeps the allocation on-demand.
Four successful Spot rows exceeded eight minutes because the planner underpriced their work. The observed forecast-to-allocation gap reached 307 seconds. Use a 320-second market-only reserve to prefer on-demand for those rows while preserving packing, workers, deadlines and interruption policy.
Host-hook session writes schedule maintenance after returning, leaving agent databases open when the fixture deletes its state directory. Drain the existing session cleanup owner and close those handles before removal; extract the fixture to preserve the contract file limit.
Resident row refresh can consume synchronous render slices while an exact worker read is pending. Capture one existing exact completion and yield before admitting the next resident slice, preserving readiness checks and caller-owned errors without letting later arrivals starve bulk progress. A controlled real-worker regression fails on the original scheduling and proves both exact response priority and bounded resident progress. The historical CI latency excess remains unproven as a whole; the unchanged latency budget still gates the qualification campaign.
Waiting for an exact caller after its database facts were accepted blocked resident materialization of its pinned archive suffix. Use the oldest row's existing dirty, cold and accepted-facts state to wait only for facts that still need to arrive, preserving the supported bulk completion path. The handoff fixture also treated fake timer admission as asynchronous refresh completion. Join the actual suffix render before asserting its order, keeping the later exact request held and every assertion intact.
Matched whole-job observations show six serial workloads exceeding the 25% RunsOn slowdown threshold in both samples. Recognize their co-resident owners after packing and retain all execution policies. Constrained native fits support five 8-class candidates; the SDK declaration and planner pair needs the 16-class for memory headroom. Actual-label qualification follows. The cron comparator lacked its peer's admission priority and appeared last in a 118-row matrix. Admit both providers together without changing row payloads, other ordering, the cap, worker limits, or deadlines. Preserve the unexplained residual scheduling delay in the evidence.
Main repacked the unified-declaration envelope from the measured 80-file workload to 47 files with 37 different companions. Its earlier constrained fit cannot qualify this new job on the 8-class runner. Preserve the original 32-class allocation while retaining its execution contract and row count. Record the expired workload bindings and current planner inventory separately from historical fit evidence. The remaining guarded retention shapes no longer match the current broad plans.
Boot and Talk relay cleanup can reach the main-thread-only shared-state broker after an unsettled native worker exit. Their ordinary Gateway thread pool cannot own that cleanup authority. Classify both files with the existing database-worker fork owner and pin the routing boundary in the existing regression table. Keep production admission guards, assertions and deadlines unchanged. The mapping regression fails on the old owner; the native exit trigger was not reproduced by either valid original-child control.
The delivery-completion fixture removes its private store while background session maintenance can still own database resources. Native observation found writable SQLite descriptors at every old removal boundary; the canonical scoped drain closes them before removal. Await the existing session cleanup owner before the unchanged zero-retry fs.rm. Keep assertions, mocks, environment and timeouts unchanged. The historical ENOTEMPTY failure was not reproduced by the original row replay.
Bind current row counts and retained compiler capacity to the rebased inventory. Record the new baseline-ratchet prerequisite so older workflow walls and small-host probes are not presented as current qualification.
The obsolete-only scenario inherited unrelated drift when main added native strings before the generated-resource refresh. Seed its existing in-memory resource overrides from the canonical catalog, then inject the obsolete row. Keep the strict validator, all negative assertions, and production data intact.
Publish/all manifest writing now seals qualified npm inputs, but the workflow command fixtures still cleared the required token and artifact facts. Supply rejecting synthetic GitHub, archive, SDK, and npm inputs through that real writer boundary and assert its sealed result. Restore the documented stable/full blocking performance policy shared by fresh manifest facts and failure classification. Keep beta advisory behavior, reused manifest evidence, artifact integrity checks, and waiver authority.
Runtime-publication regressions pushed the failover classification file past the existing 1,000-line lint limit. Move its eight unchanged remediation-hint cases into the existing diagnostics owner, with the required command import. Keep every assertion and both files in their existing CI child. Production behavior, test inventory, worker policy, deadlines, and lint limits stay intact.
Adopt the canonical fixture from #157397. Build the synthetic obsolete rows directly from generated catalog entries while preserving strict validation, live-reference checks, and cleanup.
Adopt the canonical grouping from #157406. Keep coordination-error cases in the existing diagnostics owner and retain all 92 case bodies, assertions, parameter tables, describe paths, and provider-hook isolation unchanged.
The extracted PR wrapper reaches openclaw-state.worker, whose eager plugin state contract import was absent from the closed source inventory. Native main-refresh checks then exited with a missing-module error before their intended boundary. Carry Vincent Koc's one-entry fix from 9a1a6f7 (#157394). The eager closure, native provisioning and main-refresh owner tests pass 106 cases with one platform skip; changed checks and P2 review pass.
The copied runner fixtures reached runtime-publication admission before spawning their fake child. Unlike the original-wrapper controls, they did not register the existing source-runner service fixture, so host service state could reject the command and surface as a child.pid readiness failure. Install the existing fixture in the child preload before applying prepared module mappings. Keep real runner behavior, assertions and deadlines intact. The matching cold replay changes from three failures to five passing cases. Native proof passes 20 standalone runs at eight workers, three exact 80-file CI children at two workers, and 71 owner/sibling cases. Changed checks, formatting and P2 review pass.
Successful children in run 36038423993 exposed serial work hidden by old packing estimates. Two Gateway children exceeded 370 seconds before their siblings, two other serial cohorts exceeded 500 seconds against 360-second forecasts, and a retained compiler cohort was admitted late at 48 seconds. Apply exact provider and child observations in the existing placement owner after routing. Separate intact oversized children and complete measured serial cohorts, keep conservative sibling forecasts and worker policies, and correct the compiler admission cost. Exact extension selector costs also separate the measured slow plugin pair under its existing envelope. RunsOn broad PR plans grow from 118 to 123 rows within the unchanged cap. All push plans and complete child execution contracts remain unchanged. Regressions fail on the original packing; final focused checks, all eight profile replays, changed checks, formatting and P2 review pass. Native exact-head qualification remains required for wall time and cost claims.
The diversified memory32 request still concentrated all 17 Spot allocations on one family in run 36050763810. Four workers were explicitly interrupted, including a short job whose forecast could not predict capacity loss. Select on-demand for this class while preserving family and AZ eligibility, worker limits, remaining Spot fallback and disabled interruption retries. The intact tooling cohort completed twice on two-core Blacksmith allocations in 502 and 557 seconds while retaining a 461-second forecast. Consume its complete source-cohort price before the existing long-tooling route, moving one row to general16 with a 634-second admission forecast. No shards, workers, thresholds or deadlines are added. Both new regressions fail on the prior head. Nineteen focused checks, all eight profile comparisons, type-aware lint, full workflow and changed checks, formatting and P2 review pass. Native AWS performance and final exact-head qualification remain to be measured; no Spot reliability or completed-run savings are claimed.
The branch's manifest repair overlaps the producer-only fix in #157486. Use the landed producer declarations and advisory execution policy, while preserving publication gates and explicit waivers. Remove the branch-only execution-policy tightening and align its state, summary and docs owners. Keep the rejecting synthetic artifact fixture and extend its writer table to the full profile. The focused integration passes 631 cases; source compositions, changed checks, formatting and P2 review pass. Prior native fixture proof remains bound to its tested source and is not repeated.
Ordinary CI 36064878896 passed its test jobs, but the baseline prerequisite waited 568 seconds for a hosted runner and the final gate waited another 300 seconds. The workflow took 33m40s. Reuse the existing trusted first-attempt preflight admission for Blacksmith16 baseline ratchets and Blacksmith4 result aggregation. The baseline class supplies four actual CPUs, matching its hosted sample. Missing or failed preflight keeps the existing fallback. Count the actual hosted jobs while reserving the former control slots for admission, preserving every optional offload decision. Retain runner-group wrappers, permissions, checks and deadlines.
The paid-control cases and a new Android budget table on main touched the same test insertion and manifest call, blocking the PR merge ref. Move the unchanged control cases within their existing describe and account for the two reserved slots in the local reference base. This preserves the assertions and emitted row count while leaving the manifest call compatible with main's Android selection argument.
Reconcile stdin callback, pipe error, and native exit through one transport failure operation so a startup EPIPE cannot replace its diagnostic. Preserve live-child pipe errors, protocol failures, abort identity, and immediate credential disposal without changing retirement budgets.
Deleting a presumed loose object can fail before frozen-target admission when Git stores the blob in a pack. Commit a tree referencing an absent blob and bind that commit to the fixture, preserving working bytes and proving real rejection without hydration.
Contributor
Author
|
Closing this unqualified runner-capacity experiment. The recorded samples did not satisfy the whole-run time and reliability gates, and the branch combines routing changes with separate runtime and fixture repairs that should be evaluated independently. Closed in a sweep of stale maintainer PRs (no activity for 7+ days). Reopen with a current reproduction if this still matters. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Related: #155403. Head:
bc5cdbe0a7dd6607b005c912d93ac74092ea993b.Problem and behavior
Blacksmith's requested 32/16 classes deliver eight/four CPUs; measured workloads use roughly two to five busy cores. This opt-in RunsOn profile moves suitable complete workloads to appropriately sized AWS allocations, retains measured Blacksmith advantages, and splits the observed 1,142-second CLI tail through the existing packing owner. Production remains
hybrid; this PR requests neither a backend switch nor auto-merge.AWS requests offer M8azn/M8a/C8a/M7a/C7a alternatives across every AZ exposed by the existing us-east-1 stack, preserving memory floors, on-demand capacity fallback and
retry=false. Actual instance/market/CPU/memory/AZ receipts determine the comparison; no unmeasured single-thread speed guarantee is claimed for older alternatives. Unknown or long work uses on-demand. Memory32 also uses on-demand after four explicit interruptions among 17 diversified Spot allocations in run36050763810. Every AWS row currently emitted by this candidate is on-demand, so the campaign does not qualify Spot reliability.A positive forecast plus the existing 320-second market reserve must fit within 480 seconds for Spot. This is admission policy, not a raised execution deadline or a fitted interruption hazard. The original 19/36 pooled interruption rate mixes correlated AMD losses with unaffected pools. Captured-rate break-even probabilities were 64.26% for repeated full losses and 81.97% for a hypothetical eight-minute attempt followed by on-demand replacement at the observed 189.84-second mean lost allocation. No interruption retry is enabled.
The intact 134-file/three-child tooling cohort uses its two native observations before the existing long-tooling route, selecting general16/on-demand with a 634-second forecast. Packing preserves complete inventories, builds, worker policies and the 90/70/130 caps. The RunsOn Node plans contain 123 PR rows and 65 main-shaped rows. Hosted lint/type routing stays unchanged.
Critical-path gates
Ordinary merge CI
36064878896passed all 174 executed jobs but took 2,020 seconds raw / 2,019 seconds to ci-gate. Baseline job107852905814waited 568 seconds and ran 187; final gate107860989914waited 300 seconds and ran four. These waits motivate the narrow paid-control route: existing admitted first-attempt hybrid/RunsOn contexts use Blacksmith16 for the baseline prerequisite (four actual CPUs, matching the hosted sample) and Blacksmith4 for the Bash-only final gate, adopting R4 #156129's control-job direction.One preflight decision owns routing and hosted counts. The existing admission reservation preserves every optional hosted offload decision. Checks, dependencies, permissions, release runner-group wrappers, deadlines and protected lint/type stripes remain intact. Missing/failed preflight, untrusted contexts, GitHub override, retries and ordinary manual dispatches retain their existing behavior. Native timing from the following ordinary merge run is recorded below.
Qualification
First exact-head PR sample failed: run
36072751716, attempt one atcfdd4c1446b3870463e8493f400d655c779136f6, finished in 1,280 seconds raw and through ci-gate. It passed 176 jobs, skipped 13 and failed three owner rows plus the aggregate gate. Blacksmith consumed 387.433333 machine-minutes / $9.862800; AWS reference compute was $2.896327; combined $12.759127. All 62 strict AWS allocation receipts, source receipts and market checks validate; all allocations were on-demand across two AZs with zero interruptions. Failed child coverage is incomplete. No qualification or savings credit is claimed, and no later slot or rerun has been dispatched.The three failures are a retained Blacksmith CLI transport
EPIPErace (107879799307), a tooling fixture assuming a committed Git object is loose (107879798618), and the previously bounded compiler cancellation-noise mismatch (107879798821). The transport and Git-fixture repairs are now proven below, without weakened assertions or deadlines. Hosted preflight waited 384 seconds and ran for 103; the existing live permission/PR checks must precede paid allocation, so that admission policy remains intact. Four green exact-head samples within 900 seconds and successful same-head hybrid controls are still required before the 40% main / 30% PR Blacksmith-minute targets can be claimed.The preceding
c4b2f4c787adcandidate's PR sample36050763810failed at 885 seconds: four explicit Spot interruptions, two unresolved shutdown cancellations, a Blacksmith ratchet-fixture failure and the failed aggregate gate. All 61 allocation/source receipts validated, but interrupted work lacks completed coverage. Blacksmith was 421.783333 machine-minutes / $10.634400; AWS reference compute was $2.337276; total $12.971676. This failed sample earns no qualification or saving credit and was not rerun.Ordinary merge
36064878896used 755.716667 Blacksmith machine-minutes / $29.241067 and no AWS. Its 86-selector/two-child tooling tail took 822 seconds; this differs from the frozen candidate's 134-file/three-child cohort and is not a direct speed comparison or same-head control. Reference prices exclude storage, network, control-plane, teardown and invoice reconciliation; GitHub-hosted compute is treated as free.Owner repairs and proof
The routing campaign exposed fixture and ownership defects. Completed proof is retained against its exact source; changed integration inputs receive current-head CI rather than duplicate stress campaigns.
registerSourceRunnerServiceFixturelet runtime-publication admission exit before the fake child. Native proof: 20/20 standalone runs, three exact 80-file/2,414-case children (212.26/217.96/182.47s), 71 siblings. Same-platform cold control was three failures/two passes, then five passes after repair. Run36044826845.The historical UI RPC timeout (
35859004781/107174968410) remains unexplained and its spec stays in a disjoint Blacksmith row. The matched five-case native pair passed on both providers with identical selectors/workers/runtime: complete jobs 150→94 seconds, UI steps 30→23 seconds. Those successes do not prove an environment cause. The Gateway retained-history pilot passed on RunsOn (35887809735/107273123947), 9.132 seconds for the case and 424 seconds for its job, without an NVMe/cgroup workaround.Two later failures received bounded diagnosis without speculative repairs: the original 112-file compiler-ratchet replay passed 3,248 cases but emitted cancellation noise; the exact CLI mismatch remained unreproduced. The ordinary merge UI deadline failure's original 79-file/375-case shard passed, and a focused trace captured successful +3ms request/+4ms reply ordering inside the existing 100ms advance. No provider-cause or flake-fix claim is made. All diagnostic leases are stopped.
Measured one-worker fixture costs include UI session creation 3.848s, host-hook 21.672s, delivery completion 16.054s median in its endurance policy, Gateway worker/accepted files 29.954/43.650s, lifecycle 31.60s, Android 8.983s and release artifact contract 19.072s. Existing native child timings are retained above; current CI provides candidate integration/performance evidence.
Final control-gate proof: six original runner regressions reproduced; 29 admission cases passed; 741 owner/sibling cases passed with 12 existing skips. P2 found stale preflight outputs could retain paid routing on partial reruns. The selectors now also require the current attempt to be one; retained-output regressions failed before the correction and all 49 focused cases pass afterward (11.54 seconds). Final P2 review is scoped-clean with no actionable P0–P2 findings. The actual changed-file gate, owning typecheck, typed lint/canary, full workflow validation, formatting and whitespace checks pass. Thirty other job definitions and both control jobs’ non-routing fields are unchanged.
The test-only main overlap is resolved without importing Android behavior: the control cases moved byte-for-byte within their existing describe, and equivalent reservation arithmetic preserves main’s manifest-call extension. Exact three-way merge proof is clean, all 29 affected cases pass (13.47 seconds), the actual changed-file gate passes, and its separate P2 review is scoped-clean. Production, dependency and routing inputs are unchanged by that reconciliation.
Ordinary merge run
36071199008finished failed in 925 seconds: 156 successes, 18 failures and 14 skips. Sixteen checks inherited main’s staleUpdateStepResultimport, independently repaired by #157659 (b5cfc8dd9221); that importer is absent from the frozen candidate. The separatequestion-guestUI fixture timeout is repaired on main as229ab8a1fe957d3d4822309a39ddf9c7ec616395; the aggregate gate also failed. The paid baseline waited eight seconds and ran 92; the final gate waited six and ran six. Their combined reference compute was $0.049867. The longest remaining path was hosted core lint (444 seconds waiting, 367 running), whose routing is unchanged. Total ordinary Blacksmith usage was 712.183333 machine-minutes / $28.254133, with zero AWS. Failed, different-merge evidence earns no qualification, control or savings credit.The next candidate
bc5cdbe0a7dd6607b005c912d93ac74092ea993brepairs two of the three failed frozen-owner rows. The failedcfdd4c1sample is retained unchanged; no unchanged-head rerun or later qualification slot was dispatched.36075111400.36077041209. The historical runner's exact storage mechanism remains unknown.Both owner patches have scoped-clean P2 review. The combined actual changed-file gate, formatting and whitespace checks pass. All eight saved Node profiles, 43 plugin payloads, workers, coverage and market requests are identical to the prior candidate; this is topology proof, not a runtime claim. No lockfile, deadline, original assertion, backend setting or protected hosted route changed.
Compiler cancellation noise remains unresolved. The upstream native-server shutdown fix in microsoft/TypeScript#64276 is merged, but the currently published stable dependency is still 7.0.2. No stderr filter, dependency patch/upgrade, private transport access, or broadened shutdown budget is included. Hosted qualification admission still performs its existing live authorization checks before paid allocation, and R4 #156129 remains open. Qualification and the savings targets remain unproven.
The main-only guest-question fixture failure was deterministic: manual admitted authority lacked the live role-assignment reader added by main, so question registration threw before responding and its deferred never settled. The existing Gateway capture owner now constructs the fixture authority and cleanup releases it. The original native shard failed; the unchanged test passes 20 standalone runs, three original 78-file/353-case children (370.956/356.626/371.884 seconds), and 57 sibling cases. P2 review is scoped-clean. Fresh integration also passes the focused E2E, owning UI type graph, actual changed-file gate and source checks. The one-file repair landed directly on main as
229ab8a1fe957d3d4822309a39ddf9c7ec616395, Main push CI36080144666completed successfully with onlysecurity-fastexecuted; the current hourly-CI policy skipped preflight, tests and ci-gate. That push is not UI execution proof; the native campaign/current-main integration above supply the executed evidence. It does not alter the frozen S2 source or routing.Ordinary merge run
36079187703atbc5cdbecompleted in 802 seconds raw and to ci-gate: 173 success, three failures and 14 skips. Blacksmith consumed 686.116667 machine-minutes / $27.272800 reference compute; no AWS. Paid baseline waited six seconds and ran 84; final gate waited 30 and ran six. The final compute row took 594 seconds. The run's two owner failures were the inherited main guest-question fixture (now landed above) and update/restart's VM fixture: main added a mixed default/named builtin import and real filesystem authority/removal dependencies that the fixture did not supply. The latter fixture repair is now proven and landed below without changing production or its 19 assertions. Aggregate gate failed. This ordinary merge run meets the wall bound but is red and is neither an exact-head qualification sample nor a valid hybrid control.The inherited update/restart VM fixture repair landed directly on main as
691d2e5cd48c0ddda650123b8b9ee0089fa39004. The fixture now exposes real builtin namespaces and loads the actual pinned filesystem authority/removal closure; production, all 19 native cases, assertions and deadlines are unchanged. Native Linux before/after: 16/19 pass with three default-export link failures → 19/19 pass. Twenty owner runs pass at eight workers; three saved 48-file children pass 1,500 tests each at two workers (219.595/214.022/215.836 seconds); four siblings pass 137 cases with nine existing skips. One-worker file wall is 2.559 seconds. These main-source children omit only S2's three additional planner cases, recorded separately from the failing merge's 1,503-case count. P2, typed lint, formatting and final source/runtime/lock guards pass. Fresh-main Knip input changes received a new actual changed-file gate and 39 passing dead-code contract cases, without repeating the native campaign. All associated test leases are stopped.Qualification remains incomplete: zero green qualifying samples on
bc5cdbe, no successful same-head hybrid controls, and no claim that the Blacksmith-minute targets are achieved. The compiler shutdown dependency repair awaits explicit approval under repository policy; no dependency, stderr assertion, hosted lint/type route, or production backend was changed. R4 remains open. The failed samples are retained, and no unchanged CI run was retried.The second main push run
36082149605also completed successfully with onlysecurity-fastexecuted and 33 jobs skipped under the hourly policy. Neither main push supplies test-execution or routing-qualification credit; the recorded native campaigns are the executed owner proof.