From f77ac48534766eb1433963d61b8a4f4be4a38c89 Mon Sep 17 00:00:00 2001 From: sammiller Date: Mon, 26 Oct 2026 05:06:14 +0000 Subject: [PATCH] docs: record MR-18 scoped memory evidence and stop handoff Co-Authored-By: OpenAI Codex --- ledger/ledger.jsonl | 2 + maintenance/2026-10-24-repair-list.md | 2 +- maintenance/2026-10-25-handoff.md | 2 +- maintenance/2026-10-26-handoff.md | 39 +++++--- maintenance/2026-10-26-mr18-validation.md | 112 ++++++++++++++++++++++ 5 files changed, 142 insertions(+), 15 deletions(-) create mode 100644 maintenance/2026-10-26-mr18-validation.md diff --git a/ledger/ledger.jsonl b/ledger/ledger.jsonl index 0d37faa..dfefb45 100644 --- a/ledger/ledger.jsonl +++ b/ledger/ledger.jsonl @@ -148,3 +148,5 @@ {"at": "2026-10-26T02:12:43Z", "task": "AF-NODE-02", "kind": "note", "text": "MR-14 fixed: Design#34 merge 3df9db7946e9af6fd1e43f97216adfc44fd55507 before ArcForges#56 merge 5fa2ad5fe77bef791b96db11129b3851fe55c940, reviewed head 0d261bbe1d48f3c1e9848c15f32a26132cf030da. Existing secure_delete plus manual post-commit/replay WAL TRUNCATE and periodic zero-new-row retry; busy/failed manual checkpoint returns existing INTERNAL/internal, effect HAPPENED, no rollback/dedup rewrite. Real before 6 failures per native host, after all 6 own-platform regressions pass; full Store Windows95/97 (2 platform skips), Linux94/97 (3), affected AOT zero warnings/errors. Final complete Windows13/1148s and Linux16/1364s passed at exact head; Design Windows3/15s Linux3/34s. Evidence maintenance/2026-10-26-mr14-validation.md. Direct work/review by OpenAI Codex alone. No new schema migration, restricted credential operation or task activation. MR-16 remains blocked; MR-07 and MR-11 deferred/unresolved. Owner's 2026-10-26 instruction: finish MR-14 then stop for context adjustment to 872000; MR-15 remains next and unstarted. Current handoff maintenance/2026-10-26-handoff.md; 35 disposed,67 open,2 deferred,1 blocked of105. Design#35 merge 2d9995e78c4aa0f6327371ab0b5639857718da5c resolved the common idempotency-summary review finding before source merge; native full Windows3/26s Linux3/45s passed, no source/profile input change."} {"at": "2026-10-26T02:27:22Z", "task": "AF-NODE-02", "kind": "note", "text": "MR-15 audited, unresolved/deferred AR-22, not fixed or validated: ArcForges#12 original Windows/WSL-concurrent CrashConsistencyTests(420) all20 unavailable has no retained unsealing reason; historical Windows selected real Credential Manager, whereas 6978b0c moved writer/reopen to file-backed isolated keys. Read-only Windows audit1940 text logs/11920238bytes across88 run directories/86 summaries (70win,16legacy host-unspecified;75success,11other) found only6 explicit isolated case passes, no original failure/cause; no read errors or size skips. Audit ba16cbe71de88665228e807460632a52c1915591f5d42fe36d28e6cd96f8c0c4. Existing native source profiles and isolated Store passes do not prove the old real-store path. Evidence maintenance/2026-10-26-mr15-observation.md, owning AF-NODE-02 AR-22 closure requires actual nonsecret causal evidence, no new OS-store operation/stress/cleanup authorized. MR-16 remains blocked; no runtime/Design patch, assertion relaxation, real-store test, migration or task-state activation. Owner resumed on2026-10-26 and requests remaining counts after each item:35 disposed,66open,3deferred,1blocked of105;70 unresolved total. Next MR-17, skip blocked MR-16. Direct OpenAI Codex work/review only."} {"at": "2026-10-26T03:23:31Z", "task": "AF-EXEC-01", "kind": "note", "text": "MR-17 fixed: Design#36 merge9da4127003a90842ddafeb03a6f07e7c5877f34c before ArcForges#57 merge852d23d8b4daca78cef509cb7a6925b18795b6b1, reviewed head92235c31a30d7434a3abe78c09083c3bd507bf9e/tree8013bd2effc7df38b2c14000b073f2a5f643ccd0. Holder retains job_memory from own durable peak above actual enabled notification threshold after released-memory/query sequence; nonzero-exit condition and original hard fallback remain, not proof of allocation refusal. Initial4/4 and5/5 diagnostic batches retained; discriminating before9ec4939 two recovery reason failures plus3 controls, no skips,13.413s. Final author Windows new5/5 JIT+5/5 AOT holder, TaskControl15/15,Job8/8,Repo72/72,HostAOT0warnings/errors; LinuxRepo72/72,pureformat14/14,HostAOT0warnings/errors. Full final pushed-head Windows13/1163s and Linux16/1364s passed, all results/statuses read; Design Windows3/12s Linux3/45s. Evidence maintenance/2026-10-26-mr17-validation.md. Direct OpenAI Codex work/review alone; no new schema migration, real credential operation, altered assertions/steps, old-head pass or later task activation. MR-16 blocked;MR-07/11/15 deferred. Counts36 disposed,65open,3deferred,1blocked of105;69 unresolved total. Report counts and continue MR-18 under owner's resumed authorization."} +{"at": "2026-10-26T05:05:10Z", "task": "AF-EXEC-01", "kind": "note", "text": "MR-18 fixed: Design#37 merge973934c658f62967be4b6cc40d0526d877ee4383 before ArcForges#58 mergea4c1b2bf1b8afd0112fb490aa590c3ac8199ac96, final reviewed head15d6717144468ce0be615e8538f858d29b0fed7d/treefed15ef1b4251c08d8bd48afcfe32b2530570e06. Own-job notification/peak or hard-limit peak evidence gates memory packets; qualifier leaves allocation refusal unproven; MR-17 comparisons preserved/shared. Before detail3 failures/2controls plus actual kernel hard packet10 on real shared port wrongly own hit=True (own peak117784576<268435456), not the three default-forwarding prerequisite timeouts. fa6f836 authors memory5+1 JIT and5+1 AOT holder passed; final delta only fairness isolation/comment, no src/tools/eng runtime change. fa6f836 Windows full CI failed timed share65.3948% despite6:6 starts; standalone50.5%, new-head50.9%; existing exclusive policy applied without changing assertions/timeouts/loads/algorithm. Final complete Windows13/1140s and Linux16/1382s passed at final head; all results/statuses read; Design Windows3/12s Linux3/44s. Partial Linux launcher/controller run and old-head results retained, never final status substitutes. Evidence maintenance/2026-10-26-mr18-validation.md. Direct OpenAI Codex work/review alone. No new schema/RPC/key/credential operation or task activation. MR-16 blocked;MR-07/11/15 deferred. Owner says finish MR-18 then STOP, MR-19 unstarted. Counts37 disposed,64open,3deferred,1blocked of105;68 unresolved total."} +{"at": "2026-10-26T05:05:10Z", "task": "AF-SCHED-01", "kind": "note", "text": "MR-18 required-gate fixture follow-up (not a scheduler algorithm change): TwoSessionsShareAdmissionsEquallyOnWindows failed fa6f836 parallel CI at65.3948% running slot-time share while starts6:6. Source records include launch/finalization timing for300ms holds, so the prior independence comment was false and both OS wrappers violated existing timing-sensitive exclusive policy. One unchanged fa6f836 standalone check passed50.5%; final15d6717 standalone passed50.9%,6:6. Add isolation=exclusive and DisableParallelization to both wrappers; retain40-60%,5-7 starts,12tasks/session,300ms holds,180s timeout and all assertions; CI+exclusive retains coverage. Final source#58 mergea4c1b2bf1b8afd0112fb490aa590c3ac8199ac96, full Windows13/1140s/Linux16/1382s passed at15d6717144468ce0be615e8538f858d29b0fed7d. No new AF-SCHED-01 feature/scope/state, no weakened gate or old-head status reuse. Owner stop after MR-18 remains controlling. Evidence maintenance/2026-10-26-mr18-validation.md."} diff --git a/maintenance/2026-10-24-repair-list.md b/maintenance/2026-10-24-repair-list.md index b7b3de2..7ee0826 100644 --- a/maintenance/2026-10-24-repair-list.md +++ b/maintenance/2026-10-24-repair-list.md @@ -60,7 +60,7 @@ managed tasks, no option that disables protection to enable elevation. The stop | ID | Issue | Ledger | Repos | Evidence | Proposed disposition | Status | |---|---|---|---|---|---|---| | MR-17 | Windows: an attempt that crossed `mem_ceiling` while the node was down is classified EXITED, not MEMORY_LIMIT (the exit record's violation query reads the job state at exit, empty once memory is released). | 39 (2); ArcForges#12 | Design, ArcForges | ArcForges#12 open question (3 workstation runs) | fix: the holder records the job's peak against the notification limit; Design 04 §7 states the node-gap rule | fixed: Design#36 (merge9da4127; native full Windows3/12s,Linux3/45s) before ArcForges#57 (merge852d23d, reviewed head92235c3; native full Windows13/1163s,Linux16/1364s); durable peak compared with actual enabled notification threshold preserves job_memory after post-release queries; nonzero-exit guard/hard fallback retained, no allocation-refusal guarantee; real before2 failures/3 controls, after JIT5/5 and AOT holder5/5; [validation](2026-10-26-mr17-validation.md); ledger AF-EXEC-01 note | -| MR-18 | MEMORY_LIMIT is recorded for a non-zero exit after the commit notification at `mem_ceiling` with no allocation refused, and when a workload's own nested job reports a limit. | 55 (1) | Design, ArcForges | *(confirm in the classification code)* | fix: classification follows the specified evidence; Design states it | open | +| MR-18 | MEMORY_LIMIT is recorded for a non-zero exit after the commit notification at `mem_ceiling` with no allocation refused, and when a workload's own nested job reports a limit. | 55 (1) | Design, ArcForges | `JobObject.HasOwnMemoryLimitEvidence`, `WindowsAttemptHandle.OnJobMessage`, `ExecutionSupervisor.Outcome`; real shared-port packet regression and own-job detail cases | fix: classification follows the specified evidence; Design states it | fixed: Design#37 (merge973934c; native full Windows3/12s,Linux3/44s) before ArcForges#58 (mergea4c1b2b, reviewed final head15d6717; native full Windows13/1140s,Linux16/1382s); own-job evidence required for packet attribution, detail explicitly leaves allocation refusal unproven; before Core3 failures/2 controls plus actual shared-port hard-packet failure, after6 JIT+6 AOT-holder checks pass; default-forwarding prerequisite timeouts not counted as defect/pass; failed fa6f836 Windows timing-share gate diagnosed and existing exclusive policy applied with all bounds unchanged; [validation](2026-10-26-mr18-validation.md); ledger AF-EXEC-01 and AF-SCHED-01 notes; owner requests stop here,MR-19 unstarted | | MR-19 | `windows_status` labels every code at or above 0xC0000000 as NTSTATUS, including customer-bit codes (0xE0434352, 0xE0000008). | 55 (2), 63 | ArcForges, Design | `ExecutionSupervisor.cs` 1994 | fix: customer-bit codes are application codes (not NTSTATUS); Design and node agree | open | | MR-20 | OOM_KILLED says "killed at its memory limit" for four causes (its own MemoryMax, the slice's, a machine OOM, systemd-oomd). | 56 (1) | Design, ArcForges | *(confirm)* | fix: classify from `memory.events.local` of the attempt and the slice and systemd-oomd's attribute; detail names the cause | open | | MR-21 | The Critical pressure level reacts to any OOM kill, including an attempt's kill at its own ceiling (the machine-wide count). | 58 (1), 57 (5) | Design, ArcForges | `PressureEvaluator` uses the machine-wide `OomKills` | fix: kills inside an attempt at its own limit do not raise Critical; Design 03 §5.4 states it | open | diff --git a/maintenance/2026-10-25-handoff.md b/maintenance/2026-10-25-handoff.md index 6c6e287..ab55457 100644 --- a/maintenance/2026-10-25-handoff.md +++ b/maintenance/2026-10-25-handoff.md @@ -1,4 +1,4 @@ -> Superseded operational state and continuation instruction: see [2026-10-26 handoff](2026-10-26-handoff.md). The owner resumed maintenance on 2026-10-26 and requests remaining counts after every item. Refresh refs and host state; MR-16 remains blocked. +> Superseded operational state and continuation instruction: see [2026-10-26 handoff](2026-10-26-handoff.md). The latest owner instruction is to finish MR-18 and STOP; MR-19 is unstarted. Refresh all refs/host state on any later explicit resume. MR-16 remains blocked. # Maintenance pass: session handoff, 2026-10-25 diff --git a/maintenance/2026-10-26-handoff.md b/maintenance/2026-10-26-handoff.md index b5f5fe2..5c061b4 100644 --- a/maintenance/2026-10-26-handoff.md +++ b/maintenance/2026-10-26-handoff.md @@ -2,15 +2,15 @@ ## Latest owner instruction -On 2026-10-26 the owner resumed the authorized maintenance list after adjusting context and asked for the remaining -counts after every item. Continue serially in maintenance priority order, completing each item's review, required native checks, merge, -maintenance list and CLI ledger record before reporting the new counts and automatically continuing. -This supersedes the temporary finish-MR-14-and-stop instruction. Work alone; do not start AF-REC-01 or later features. +On 2026-10-26 the owner asked to finish the current item MR-18, including validation, merge and records, then STOP. +MR-18 is complete. Do not start MR-19 or continue automatically. A later explicit owner resume may continue the authorized +maintenance list from MR-19. Report the remaining counts at this stopping point. Do not start AF-REC-01 or later features. +This supersedes the earlier resumed continuous-execution instruction and keeps all existing safety/scope boundaries. This supersedes the resumption instruction and stale operational state in the [2026-10-25 handoff](2026-10-25-handoff.md). Scope/order and current dispositions are in the [repair list](2026-10-24-repair-list.md); read the append-only ledger too. -## Completed MR-14 and MR-17; current maintenance position +## Completed repairs and current stopping point - MR-14: [Design#34](https://github.com/ArcForges/ArcForges-Design/pull/34), merge 3df9db7946e9af6fd1e43f97216adfc44fd55507, precedes [ArcForges#56](https://github.com/ArcForges/ArcForges/pull/56), merge 5fa2ad5fe77bef791b96db11129b3851fe55c940. @@ -27,7 +27,7 @@ Scope/order and current dispositions are in the [repair list](2026-10-24-repair- - Additional [Design#35](https://github.com/ArcForges/ArcForges-Design/pull/35), merge 2d9995e78c4aa0f6327371ab0b5639857718da5c, aligned AD-0004 and the common replay summary before source merge; both native full legs passed (Windows 3/26 s, Linux 3/45 s). That was the Design context used by the MR-14 record; current verification fetches Design main afresh. -## Latest completed repair: MR-17 +## Prior completed repair: MR-17 [Design#36](https://github.com/ArcForges/ArcForges-Design/pull/36) merge9da4127003a90842ddafeb03a6f07e7c5877f34c precedes [source#57](https://github.com/ArcForges/ArcForges/pull/57) merge852d23d8b4daca78cef509cb7a6925b18795b6b1. @@ -37,14 +37,27 @@ native full Windows13/1163s and Linux16/1364s passed at that head. before2-fail/3-control and after JIT/AOT-holder5/5 cases. Current job notification threshold, strict peak crossing and nonzero-exit condition; no hard-allocation-refusal proof, new schema/key operation or later feature. The owning AF-EXEC-01 completion note is appended only through Plan CLI after source merge. -MR-18 correlated/nested-limit classification/detail remains next; MR-16 and AR-20/21/22 remain unchanged. +MR-18 is completed below. MR-16 and AR-20/21/22 remain unchanged. + +## Latest completed repair: MR-18 + +[Design#37](https://github.com/ArcForges/ArcForges-Design/pull/37) merge973934c658f62967be4b6cc40d0526d877ee4383 +precedes [source#58](https://github.com/ArcForges/ArcForges/pull/58) mergea4c1b2bf1b8afd0112fb490aa590c3ac8199ac96. +Final reviewed head15d6717144468ce0be615e8538f858d29b0fed7d/treefed15ef1b4251c08d8bd48afcfe32b2530570e06; +complete native Windows13/1140s,Linux16/1382s passed at that head. +[Evidence](2026-10-26-mr18-validation.md) separates actual shared-port packet attribution from three default-forwarding +delivery-prerequisite timeouts, and records the failed fa6f836 Windows fairness gate plus corrected test isolation. +All fairness bounds remain. Memory-code author fa6f836 results are distinct from final15d6717 gate results. +The initial malformed Linux launch/partial run and corrected old-head run are retained, not final gating evidence. +Own-job scope and uncertainty are explicit; no new RPC/schema/key operation or feature. CLI notes on AF-EXEC-01 and +AF-SCHED-01 follow actual source merge. MR-19 is next and unstarted; STOP under the owner's latest instruction. ## Remaining and boundaries -105 list items: 36 disposed, 65 open, 3 deferred, 1 blocked: 69 unresolved total. +105 list items: 37 disposed, 64 open, 3 deferred, 1 blocked: 68 unresolved total. MR-15 has been audited and recorded as unresolved/deferred AR-22, never repaired or passed; see [the observation](2026-10-26-mr15-observation.md). Current isolated-key passes cannot validate its historical real-store cause. -MR-18 is next eligible; MR-16 keeps its existing block. Check actual logs/processes/results before retrying a timed-out launch. +MR-19 is next eligible but unstarted; stop until the owner resumes. MR-16 keeps its existing block. Check actual logs/processes/results before retrying a timed-out launch. MR-16 stays blocked: no real Windows Credential Manager cleanup/stress or orphan credential-file operations; preserve the recovery proposal. MR-07/AR-20, MR-11/AR-21 and MR-15/AR-22 stay unresolved/deferred. Other acceptance-register limitations, @@ -58,9 +71,9 @@ Refresh GitHub PRs/statuses, git refs, all applicable AGENTS/governance and Desi Recorded SHAs and PIDs are historical evidence, never live state. Preserve all worktrees, user changes and ignored caches. Primary checkouts intentionally remain stale; use fresh origin/main and task worktrees. -Windows source worktree: ArcForges/.worktree/mr-17 (clean detached final head). -Linux source worktree: ArcForges/.worktree/mr-17 (task/mr-17, clean final head). -Plan reusable Windows worktree: Plan/.worktree/maintenance-resume-2026-10-25, branch plan/maintenance-mr17-memory-gap. -Formal logs remain under each repository's .git/arcforges-verify/logs; mr14-*/mr17-* wrapper logs are in the hosts' dev/logs or logs. +Windows source worktree: ArcForges/.worktree/mr-18 (clean detached final head). +Linux source worktree: ArcForges/.worktree/mr-18 (task/mr-18, clean final head). +Plan reusable Windows worktree: Plan/.worktree/maintenance-resume-2026-10-25, branch plan/maintenance-mr18-memory-scope. +Formal logs remain under each repository's .git/arcforges-verify/logs; mr14-*/mr17-*/mr18-* wrapper logs are in the hosts' dev/logs or logs. Never terminate the Linux user systemd manager, use pkill -f or daemon-reload. Keep formal hosts free of competing tests. Work alone under the owner's instruction; use truthful OpenAI Codex attribution and direct review. diff --git a/maintenance/2026-10-26-mr18-validation.md b/maintenance/2026-10-26-mr18-validation.md new file mode 100644 index 0000000..9ae0a51 --- /dev/null +++ b/maintenance/2026-10-26-mr18-validation.md @@ -0,0 +1,112 @@ +# MR-18: own-job memory evidence and honest limit detail, 2026-10-26 + +MR-18 is fixed. [Design#37](https://github.com/ArcForges/ArcForges-Design/pull/37) merged first as +973934c658f62967be4b6cc40d0526d877ee4383. [ArcForges#58](https://github.com/ArcForges/ArcForges/pull/58) merged as +a4c1b2bf1b8afd0112fb490aa590c3ac8199ac96, final reviewed head15d6717144468ce0be615e8538f858d29b0fed7d, +treefed15ef1b4251c08d8bd48afcfe32b2530570e06, base852d23d8b4daca78cef509cb7a6925b18795b6b1. + +## Repair and scope + +MEMORY_LIMIT is a normalization of own-job evidence plus nonzero exit, not proof of memory causation or allocation refusal. +The old detail said the attempt reached its commit limit even when only the lower notification threshold had been crossed. +The new detail explicitly names the attempt's own job and that allocation refusal is not proven; raw exit/effect rules remain. + +The old hard-memory packet branch latched MemoryLimitHit without checking the attempt's own limits. Confirm actual own-job +notification/peak or existing hard-limit peak evidence before latching, using one predicate shared with the holder so MR-17 +durable evidence keeps exactly the same comparisons. Descendant usage is included in own aggregate usage, but a descendant's +lower cap is not the attempt cap. Queued packets after disposal fail closed on the expected disposed-handle path. +No scheduler/enforcement algorithm, enum/RPC/package/schema/migration or credential operation is changed. +AD-0005/date,04 section7 and internal contract comments state evidence scope/uncertainty. + +Internal read-only packet observation and a borrowed real-port accessor support test provenance/composition. A test associates +actual jobs and observes kernel packets; it never posts a packet manually or calls the handler with simulated data. The borrowed +port remains owned by the platform. Production has no observer by default. + +## Before evidence: separate discriminating failures from delivery prerequisites + +Classification logic was unchanged in the baseline stages; read-only observation/composition seams were present. + +- ab04206f2eb70a07bdd8808d5e765957e9d5ac8c: Core node-gap/detail cases3 failures+2 controls, no skips14.663 s, + testexit2. Existing MemoryLimit reason/marker was correct but the detail lacked own-job scope and uncertainty. +- Three default/ordinary parent-forwarding experiments (mr18-before, before2, before3) timed out on the required packet-delivery + precondition. With a separate child port, the child port captured real hard packet10; raw parent trace showed process packets, + not the required hard packet. These are incomplete delivery experiments, not attribution failures, validation passes or proof + that forwarding is impossible. No retry of them was counted as a repaired-code pass. +- cc6ddbacc709043cead8ddf951a569eae093ce4d (tree7d764d4f4391630c1c6f224c56a102aa728196f8), mr18-before4: + explicit real shared-port composition with a key read from a real root process packet. Child29692 exited3762504530; + nested peak110678016 bytes, attempt peak117784576 below own notification268435456 and hard402653184. + The kernel delivered actual hard packet10 to the dispatcher and old own hit=True. Assert.False failed, no skips4.292 s; + build0 warnings/errors12.01 s, testexit2/outer wrapper0. This is the discriminating scope failure. + +The shared-port case is explicit and does not establish default-forwarding delivery on this host. No fake events, manually +posted packet, fake accounting value, asserted-unobserved success or weakened predicate/bound substitutes for actual delivery. + +## Memory-code author evidence at fa6f836 + +Headfa6f83622c8a2634b59b069dbb124e525b83ffce/tree8778dc64a3340f3f2e93c8ded1798ddcfd380e13: + +| Check | Native Windows | Native Linux | +|---|---|---| +| Format,Release build/analyzers |0 warnings/errors12.88 s |0 warnings/errors15.43 s | +| Own/detail regression | JIT5/5,no skips14.736 s; AOT holder5/5,no skips13.810 s | Windows behavior not inferred | +| Real shared-port nested packet | JIT1/1,no skips4.541 s; AOT holder1/1,no skips2.901 s | Windows behavior not inferred | +| TaskControl and JobTests |15/15 no skips76.477 s;8/8 no skips14.230 s | unchanged platform behavior | +| Repository guards |72/72,no skips3.292 s |72/72,no skips0.971 s | +| Pure format/protocol tests | full profile coverage |14/14,no skips0.597 s | +| Host Native AOT |0 warnings/errors42.37 s |0 warnings/errors47.89 s | +| Final diff/format/clean author wrapper |pass,exit0 |pass,exit0 | + +The real node is in-process for the Core tests; AOT variants use the published holder, not a new standalone native-node +nested-dispatcher test. Jobs, allocations, queries, port and packets are real; external keys are isolated. The nested test +requires an actual hard packet and offers at most4 bounded real allocation opportunities for best-effort hard messages. +A suspended child's actual process packet is observed before it allocates. Fixtures clean only their own temporary data/jobs. + +## Required-gate failure and necessary timing-isolation repair + +The first Windows full-profile attempt at fa6f836 failed win-ci: SchedulerTests.TwoSessionsShareAdmissionsEquallyOnWindows. +The first12 starts were6:6, but slot-time share65.3948% violated the unchanged40–60% bound. Core322 total:283 passed, +38 platform skips,1 failed. Summary20261026-041700-fa6f83622c8a-win is false, wrapper1,120 s; later steps did not run. +This is not a final profile pass. + +FairnessAsync uses300 ms holds but measures admission-to-finish slot time, including launch/finalization delays. Its comment +claiming independence from below-normal process startup was false; its two OS wrappers lacked exclusive isolation. +AGENTS explicitly requires timing/CPU-share-sensitive checks to use isolation=exclusive and DisableParallelization. +A single matching fa6f836 standalone diagnostic passed unchanged assertions:6:6 starts,50.5% share,10.048 s. +This supports the isolation diagnosis, not a proof about every possible source of timing variation. + +Final delta15d6717 changes only both OS fairness isolation attributes and that comment. It changes no scheduler algorithm, +300 ms hold,12 tasks/session,5–7 first-start count,40–60% share assertion or180 s timeout. CI plus exclusive retains the same +tests, and both JIT/native exclusive gates run the timed check. git diff fa6f836..15d6717 -- src tools eng is empty. +The final-head Windows delta author passed format/build0 warnings/errors25.83 s, fairness1/1/no skips12.075 s, +6:6 starts/50.9% share,Repository72/72/no skips3.127 s,diff/clean wrapper0. Prior memory-code author results above are distinct +fa6f836 evidence, never presented as final-head statuses. The final full profiles below rebuilt/checked the actual final head. + +## Linux orchestration recovery and old-head boundary + +The initial fa6f836 Linux launch used a malformed compound background command and the MCP call timed out. Logs started, +but native process/PID-file inspection found the controller gone, no summary/exit and no residual compile/test processes. +Partial logs mr18-final.out/.err and summary directory20261026-121419-fa6f83622c8a-linux remain; no pass claimed. +A corrected standalone nohup run used mr18-final2 and the exact same candidate fa6f836. After the Windows failure required +a new head, that old Linux run was allowed to finish/clean private units normally. No user systemd manager was terminated; +no pkill -f,daemon-reload,assertion/step bypass or manual status report. The corrected old-head run completed all16 steps/1349 s, +wrapper0, summary20261026-121651-fa6f83622c8a-linux. Those fa6f836 results do not gate the new commit. +Final15d6717 uses mr18-final3 on both native hosts, with no competing same-host formal build/test. + +## Final pushed-head gates, review and record + +Design head27322c407f478ac4902bd82b1f547f394641e706/treea72f8ae2ff85c4a4e9ad85f36dfa80ce23b89032: +complete native Windows3/12 s,Linux3/44 s passed, all6 results/wrappers0 and required exact-head statuses read with fresh refs. +[Design review/gate](https://github.com/ArcForges/ArcForges-Design/pull/37#issuecomment-6408024200). +Design summary IDs20261026-040704-27322c407f47-win and20261026-120657-27322c407f47-linux. + +Final source15d6717144468ce0be615e8538f858d29b0fed7d: complete native Windows13/1140 s and +Linux16/1382 s passed, all29 results/both wrappers0 and required engine-reported exact-head successes read; +owner-hosted checks/real OS-store tier skipped as expected. No old-head/partial pass or earlier author result substitutes. +Final summary IDs:Windows20261026-043314-15d671714446-win;Linux20261026-124013-15d671714446-linux. [Final source gate](https://github.com/ArcForges/ArcForges/pull/58#issuecomment-6408659480). +Author/review evidence and the timing diagnosis are linked from [source#58](https://github.com/ArcForges/ArcForges/pull/58). +Direct OpenAI Codex work/review alone under explicit user instruction; no other model/agent or independent reviewer claimed. + +After actual source merge, Plan CLI appended one AF-EXEC-01 completion note and one AF-SCHED-01 gate-fixture note. +The MR-18 row, this evidence and the stop handoff are updated; no graph/task activation or later feature. MR-16 remains blocked and MR-07/11/15 +remain unresolved/deferred. Latest owner instruction: finish MR-18 then STOP; MR-19 is next but unstarted. +Counts after this item:37 disposed,64 open,3 deferred,1 blocked of105;68 unresolved total.