docs: record final issue 15 fix-wave acceptance and limits

This commit is contained in:
阿白叔 committed 2026-11-17 13:55:20 +08:00
1 parent c1e259603a
commit 8869093eb0
2 files changed
+63 -6

No files matched your search

@@ -1,6 +1,6 @@
# Issue 15 acceptance matrix
Refs #15. The round-1 complete suite at `8818fa9` passed **2,358 tests in 451.34 seconds**, exit code 0, with zero failures, errors or skips. B9 automatic release/prequeued Versions and B10 legacy release/conditional buttons retain their actual build identities. Ordinary Verify early cancellation still retains a later loaded Forward prompt and is an open blocker. Source Qt/Windows tests, source native runs and frozen native runs are separate evidence classes.
Refs #15. The final grouped review wave is implemented at `c1e2596` (B14). The complete suite passed **2,418 tests in 594.83 seconds**, actual exit code 0, with zero failures, errors or skips. Actual B13 ordinary Verify/Stop with automatic candidate-owned cleanup and cross-owner queued Versions passed; actual B13 CBB recovery passed source and target verification. B14 preserves explicit ordinary release promotion of an existing cold observer; that narrow change has source/real-process coverage and updated frozen checks. Native cases retain B13 identities. The final scoped independent review is pending.
## Executable bindings
@@ -19,6 +19,9 @@ Each case retains its actual executable hash. Rebuilds checked all packaged modu
| B9 | a3f69de | `bfd6c122724f71a2aac4748ed01336a9949aa753f9c58aca55db574668140083` |
| B10 | 701e5cd | `a6271e4219a02123e301ed1e0ea3109e78307bf1f92574b4a5cacb70690695a4` |
| B11 | 8818fa9 | `03452e39f394c3c63113aaa701f64bacf14e817ae3b4e194965f333ecc645b98` |
| B12 | 906cfc0 | `7d7929d6621ae6c379205380e06f94eb076135343417d000018a8ab04efea8cc` |
| B13 | 5e9509e | `966e00fe57eccbdf4975c624026d39d24d4613f4174a932db248fb444a7caa9d` |
| B14 | c1e2596 | `0cd5f37b113802bcf3abd07a94d8c5b9f73d40696b1c56af66fedf39112e2d8c` |
B9 ZIP SHA-256: `4c589bae6bf7b68732cd7ebe373b2924016c71f6731fd661c1cd0fd848427438`. All 68 packaged modules and required assets/scripts/bridges/icons match the source; all 242 extracted files match the build. Its five pre-cancelled worker entry routes and common GUI/worker/owner-probe/dialog-agent smoke passed without launching Mentor. B8 is separately archived with ZIP SHA-256 `795010cc0b54b995eced859cac1513bf105092e484a44b1231955801deed879f`; B5 retains ZIP SHA-256 `176c7cda89c2a2706f969fa6dca513d2760615491aaf70c19fe7a0688abe471a`.
@@ -26,6 +29,8 @@ B10 ZIP SHA-256: `d4bd42416ea8d12a249ff1d9c8854eb76d2f81bb7ff869ff43dd92a331c647
B11 ZIP SHA-256: `34b3646b27afae22ec43a415ca19bbd32a91c0a4c941d37d6289c76cea9ca730`; the same module/resource/file checks, five cancelled entries and common smoke passed. Its only source-module change from B10 records each acknowledged Tip per cleanup intent. Immediate/delayed Tip states with/without prior Close passed focused source tests; no new actual Mentor Tip replay is claimed. The harmless real Win32 adapter probe belongs to the B10 source stage; its unchanged adapter preserves hidden MDI siblings and sends no action with an extra document.
B14 ZIP SHA-256: `a580e3e9945261761e0eb3968b4e271d9f4745ce982dea96c283af4ab06e565f`; 68 modules/resources, 242 extracted files, five cancelled entries and common smoke passed. B13 actual native cases retain their original hashes. B12 failed late ordinary cancellation remains archived with its exact identity.
## Reported issues A–H
| Issue | Source verification | Actual frozen evidence and limits |
@@ -64,10 +69,13 @@ These identifiers locate retained local evidence; raw field files, private machi
| B9 automatic cancellation and prequeued Versions | icdb-check-b05147dc18bc / verify-3abaddb3e141; native-backup-005da267ebfe | Versions queued before Stop/handoff; UI terminal 6.672 s. Automatic Tip/Forward No/one Close; 307 read-only snapshots and kernel exit. Existing observer obtained fresh library RELEASE_CONFIRMED; queued native client created 1.684 s later and catalog loaded. Zero Release clicks/manual diagnostic native actions. Original CANCELLED/safe=false retained; backend duration 12.829 s is distinct from UI timing. |
| B5-to-B9 legacy release — one confirmation | icdb-check-378044ad1122 / verify-d2acda2739c2; sessions-a5a9fe06b7c1 | Real B5 retained loaded target/Forward; old app normally exited. B9 initial inspection succeeded, one normal Confirm, zero Recheck/second confirmation/manual native actions. Forward disappeared after 3.562 s; 210 read-only snapshots show kernel exit. PCB, launcher and server RELEASED, fresh RELEASE_CONFIRMED at 22.634 s; UI agreed and Confirm disabled. Existing cold observer RELEASED; original cancellation result retained. |
| B10 I3 setup — second pending item | management-f25df6a207ca4744ab4a0c23ddabb06c | Legitimate exact-name update from the existing healthy source, auto-verify off; UPLOADED/safe, native catalog/readback. Prepared two pending items without fabricated native failure. Selection made before final refresh was cleared; disabled Verify rejected the harness click; reselection afterward succeeded. |
| B10 ordinary Stop — open defect | verify-e464474b7bab; batch-20261117-105152-c3e162 | Actual Stop 0.547 s after PCB registration, no Save evidence: first CANCELLED/safe=false, 29.766 s; second SKIPPED with no native job. Later loaded Forward remained. Exact manual No, no-modal/no-progress check and one normal Close required; sessions-4aadbacb8a93 then fresh release. Not unattended PASS; ordinary cleanup does not have diagnostic-only authority. |
| B10 ordinary Stop — historical failure | verify-e464474b7bab; batch-20261117-105152-c3e162 | Actual Stop 0.547 s after PCB registration, no Save evidence: first CANCELLED/safe=false, 29.766 s; second SKIPPED with no native job. Later loaded Forward remained. Exact manual No, no-modal/no-progress check and one normal Close required; sessions-4aadbacb8a93 then fresh release. Not unattended PASS; ordinary cleanup does not have diagnostic-only authority. |
| B10 actual Continue, no replay | button-watch-1794884270925407000; verify-f673ac48f600 | Existing summary's three-second Close elapsed before inspection. After main-window Release proof and normal dialog close, production code recreated Continue. Actual button click started only the previously skipped second item, which passed GREEN/safe. First terminal result predates the click and no first-item job was replayed. No in-process method/result injection. |
| B10 actual result Release | button-watch-1794884430968047900; verify-4f959f61fbc4 | One bounded same-scenario timing repeat caught actual result Release and opened its residual plan. Ordinary Stop again retained Forward, 32.844 s; this remains a failure. Exact manual No/normal Close, no Save/force; sessions-6bfdf0a26476 fresh release. |
| B5-to-B10 affected legacy release | icdb-check-92a3c92f6d6c / verify-9f3856153c79; sessions-82cc60b6eacb | Initial inspection succeeded; one Confirm, zero Recheck/manual native actions. Forward disappeared +6.5 s; 278 read-only samples/kernel exit. Intermediate PARTIAL then fresh RELEASE_CONFIRMED +35.387 s under the same intent; all three clients and existing observer RELEASED, UI agreed. Original CANCELLED/safe=false preserved. Entry was loaded Forward; actual Tip replay is not claimed. |
| B12 ordinary Stop — late loaded failure | verify-642171759614 | UI terminal 7.797 s; loaded Forward at 34.406 s after original cleanup deadline. Manual exact No/normal Close required. Failed-run cleanup only; original CANCELLED/safe=false preserved. |
| B13 ordinary Stop / cross-owner queue | verify-71aee0603183; native-backup-b63e7f708fd5 | Actual candidate worker No/one Close before first 30 s deadline; late kernel exit and fresh candidate/independent proofs. UI terminal 7.375 s; real Versions creation 5.220 s after candidate proof and 1.506 s after independent proof, catalog loaded. No Release/manual ordinary native action. Retained-phase No/Close branch not replayed in Mentor. |
| B13 CBB backup recovery | backup-restore-3f98efb14448; source verify-06c8149375e6; target verify-f902921f2440 | Both PASS/GREEN/safe and COMMITTED/CLOSED/EXITED/RELEASED, installed=true; old Work retained, all 366 original source files unchanged, no native intervention. |
The B2 diagnostic run supports the implemented sole-handler No sequence through observed windows and verified code; there is no separate startup button acknowledgement event. The hidden-owner-chain save branch did not recur on corrected native runs and remains covered by regression tests plus the original captured chain. The original failed diagnostic required manual cleanup and a native crash acknowledgement; that crash was **NOT REPRODUCED** in the subsequent bounded runs.
@@ -78,25 +86,45 @@ The B2 diagnostic run supports the implemented sole-handler No sequence through
| Tabs, advanced options, optional LM checkbox, runtime log toggle | Actual source UI and Qt tests | B1 actual navigation/toggles. |
| Library chooser Open/Cancel and Read CBB | Actual chooser/catalog; invalid-input tests | B1 chooser cancellation and catalog; B3 removal/reimport catalog; B4/B5 explicit reload. |
| Select visible, Clear, filter/hidden selection | Actual UI and 381/3,000-row Qt checks | B1/B2 actual controls and view changes; immutable original operation retained. |
| Verify selected / visible non-READY | Actual source native verification and Qt admission | B2 ordinary Verify; B3 recovery target verification. READY exclusion remains enforced. |
| Verify selected / visible non-READY | Actual source native verification and Qt admission | B2 ordinary Verify; B3/B13 recovery target verification; B13 actual ordinary Stop and automatic release. READY exclusion remains enforced. |
| ICDB batch confirmation No/Yes and stop | Actual source checks; Qt stop/continuation | B1/B2 confirmation cancellation and diagnostic runs; B5 acknowledged close; B8/B9 actual startup Stop and automatic queued continuation. Other cancellation timing remains conditional. |
| Manual Repair Cancel/Confirm | Actual native healthy control | B1 native diagnostics/defragmentation; not field-fault recovery proof. |
| Versions / create backup / restore Cancel/Confirm / verify / Close/reopen | Actual exact archive and location | B1 native tool operations and reopening. |
| Add / source chooser / empty input / auto-verify / logs / No/default Yes | Actual source and Qt | B1 duplicate-plus-independent Add; B3 reimport. |
| Same-name update / wrong source / reselection / No/default Yes | Actual source and Qt | B2 mismatch rejected before writing; B3 correct update. |
| Remove / hidden source chooser / No/Yes / result | Actual source and Qt | B3 REMOVED and native catalog absence before reimport. |
| CBB backup recovery / empty input / chooser Cancel / No/Yes | Actual source and Qt | B3 full source-validation/install/target-validation sequence. |
| CBB backup recovery / empty input / chooser Cancel / No/Yes | Actual source and Qt | B3 original controls/full sequence; B13 full serial metadata/source-validation/install/target-validation sequence. |
| Management stop-after-current | Two-item Qt/real admission tests verify next item skipped, including failed readback and unknown release | B3 stop during native Add; mandatory current-item readback completed. |
| Queued cancellation / active dialog close | Qt queue, close, hard-cancel and shutdown tests | B10 recovery dialog queued behind an owned diagnostic, Cancel queued clicked, UI CANCELLED/no native admission/no recovery job. Underlying diagnostic verify-96d285ffa450 completed CHECKED/safe. Earlier second management submission/active lower Close remained disabled. |
| Residual sessions / Recheck / Confirm / Close | Qt exact scope, one-intent observer promotion and real Windows proofs | B5 zero-active-session release; B9 initial inspection and one-confirmation release of a real B5 cancelled session; previous failed/multiple confirmations retained. |
| Native details / Save prompt / native Cancel / forced termination | Qt exact ownership/prompt and default-Cancel tests | Disabled with no matching active owned client. No forced native termination claimed. |
| Conditional result recovery/release/resume buttons | Qt target binding, stale results and single continuation | B10 actual result Release and Continue/no-replay receipts above. Ordinary retained startup required manual native cleanup and remains an open defect. Other fault-dependent recovery branches are not blanket passed. |
| Conditional result recovery/release/resume buttons | Qt target binding, stale results and single continuation | B10 actual result Release and Continue/no-replay receipts above. B10 ordinary retained startup required manual cleanup; B13 automatic ordinary Stop acceptance is recorded separately below. Other fault-dependent recovery branches are not blanket passed. |
| Three-second defaults / manual-vs-deadline / hidden or closed prompts | One-shot Qt timer tests | Actual management/recovery No and timed Yes; full timing competition covered in Qt. |
| Double submission / result remaining open / view changes / restart | Qt and Windows lifecycle tests | Active execute controls disabled; modeless dialogs and view changes observed; B4 no replay. |
| Open logs/evidence / shell errors | Qt dispatch and error reporting | B10 actual Open logs/Open exception clicks opened exact application logs and verify-e464474b7bab folders; complete private Explorer breadcrumbs recorded. Only the two task-created Explorer windows normally closed. No shell-error case is claimed. |
One B1 harness edit targeted the library textbox while a modeless dialog changed control indices. The immutable original-library operation still completed and the changed view was invalidated. This is retained as a harness/misuse event, not a product search failure.
## Final-wave evidence and limits
B14 EXE SHA-256: `0cd5f37b113802bcf3abd07a94d8c5b9f73d40696b1c56af66fedf39112e2d8c`. ZIP SHA-256: `a580e3e9945261761e0eb3968b4e271d9f4745ce982dea96c283af4ab06e565f`. All 68 packaged product modules/resources and all 242 extracted files match source/build. Five pre-cancelled frozen routes and common GUI/worker/owner-probe/dialog-agent smoke passed. Full-suite log SHA-256: `d9c57b7998e02154fb703f0fe2ba1f37f73dc4687d116c4b6efc0e5ad7e5c30e`; JUnit SHA-256: `4303cc123e90689103cc4135e483ac936b04a052067e952a1861ae167c7811ea`. Native clients and writer/management mutexes were idle before and after the full suite.
B13 (`5e9509e`) retains EXE `966e00fe57eccbdf4975c624026d39d24d4613f4174a932db248fb444a7caa9d` and ZIP `7a1da94e05b85873f791cde77907bdc280106768fbf84e5d89066ab8aa57a61b`. A final self-audit found that the new ordinary explicit-release branch bypassed existing cold-observer promotion. Its focused test failed before the two-line ordering correction at `c1e2596`; 110 related tests then passed in 21.96 seconds, including actual harmless worker exit, cold-observer promotion and queue proof. The in-progress B13 full suite was interrupted through its own execution session before editing, and is archived without a completion/JUnit claim. No additional native matrix was run for this narrow B14 delta.
The three review defects had retained RED evidence: actual MainWindow subprocess exit hung on stalled hash/copy/validation; real recovery callers held the unrelated native channel and could install after abandoned metadata; ordinary cancelled startup failed to clean a later loaded target. B13 related observer coverage passed 232 tests in 34.81 seconds, including an actual two-process mutex/file race between observer and explicit release plus a process restart. These race tests stub the native evidence/action boundary; they are not Mentor window tests.
B12 (`906cfc0`, EXE `7d7929d6621ae6c379205380e06f94eb076135343417d000018a8ab04efea8cc`, ZIP `bbe68005469d8b6cf87726200ef85646e442de45d164ab2ac0aec23b51d0bd65`) retained a new failure: ordinary `verify-642171759614` returned its visible cancellation terminal state in 7.797 seconds, but Forward loaded at 34.406 seconds, after the worker cleanup budget. Exact manual No/normal Close was used solely to clear that failed run. A later independent queue-helper attempt was refused with sharing violation 32 before native creation while two releasing observers raced for the coordination lease. Both failures are archived and are not relabelled as B13 passes.
B13 ordinary `verify-71aee0603183` reached the visible terminal state in 7.375 seconds. Its candidate worker acknowledged Tip at 21.359 seconds, Forward No at 27.422 seconds and one Close at 28.750 seconds; native exit followed the original cleanup deadline. Read-only sampling confirmed kernel exit. The original result remains CANCELLED/safe=false. No Release click or manual ordinary native action occurred. The helper stayed observe-only until the actual candidate produced fresh RELEASED proof, then obtained its own independent fresh proof. Its prequeued real Versions client `native-backup-b63e7f708fd5` was created 5.220 seconds after the candidate proof and 1.506 seconds after the independent proof, loaded the backup catalog and closed normally. This is **cross-owner queue acceptance**, using source-hash-pinned production queue/observer/native-worker modules; it is not survival of work in the same app after global Stop. The candidate itself performed every ordinary No/Close action.
The B13 native run issued Close within the first worker budget, so the new retained-phase No/Close branch was **NOT REPLAYED in Mentor**. That branch has source guard, real process race/restart and existing real Win32 adapter evidence. No repeated cold-start attempts were used to manufacture a slower launch.
B13 actual recovery `backup-restore-3f98efb14448` used the current pending original target and healthy independent source A. Its isolated source and installed target both returned PASS/GREEN/safe; native save/close/release completed without intervention. The old Work directory is retained and all 366 original source files are unchanged. The 27,689-file complete validation library copy progressed through the serial metadata path without a false inactivity handoff.
Separate frozen metadata-exit harness EXE SHA-256: `7490be1988d823944307510f7cfd22c85cf8570c7c25f6ee60d97f542bb72753`. All 68 packaged product modules equal final B14. A disposable I/O stall at each of hash/copy/validation exercised actual MainWindow close and interpreter exit before releasing the stall; no late dispatch occurred and partial copies were retained. This uses an acceptance-only entry, **not stall injection into the candidate EXE**. The earlier B12/B13 harnesses and all RED/setup attempts remain archived.
The original complete fault-bearing field snapshot remains unavailable (**NOT REPRODUCED**); the constructed partial copy remains **COPY_BLOCKED** and was never opened. Windows 10 and production SMB remain **NOT RUN**. The earlier first-pass B-control PENDING result and explicit retry remain separate outcomes. No merge, publication or issue closure is performed.
## Constructed copy and remaining limits
Constructed-copy experiment: **COPY_BLOCKED; inconsistency NOT REPRODUCED** (1 attempt, maximum 3). The healthy control passed normal native Verify. On a separate disposable copy, the documented native property write and stock Save actually ran; ordinary Work copying encountered Windows sharing violation 32 at the 113th attempted file, database/icdb.dat, after 0.470 seconds. The source still completed Save/Close and authoritative PASS/GREEN with released resources. The partial clone was never opened and no native Repair claim is made for it. This fault-creation instrumentation did not modify the frozen product and does not reproduce the original field incident.
@@ -2,10 +2,19 @@
Refs #15. This candidate records local Windows 11 acceptance and its limits; it is not a published release or field-acceptance approval.
The final grouped review wave is implemented at `c1e2596` (B14). The complete suite passed **2,418 tests in 594.83 seconds**, actual exit code 0, with zero failures, errors or skips. Actual B13 ordinary Verify/Stop with automatic candidate-owned cleanup and cross-owner queued Versions passed; actual B13 CBB recovery passed source and target verification. B14 preserves explicit ordinary release promotion of an existing cold observer; that narrow change has source/real-process coverage and updated frozen checks. Native cases retain B13 identities. The final scoped independent review is pending.
The PR base is main at e4875b4. The branch also includes the previously approved efb8cae three-second confirmation defaults.
The change separates native task execution, client exit, coordination-lease release, data recovery and eligibility for the next write. A completed independent item no longer inherits another project's address or a stale historical blocker. Unknown ownership still requires fresh proof; the application does not delete native address or lock files.
## Final grouped review fixes
- Ordinary pre-document cancellation now requires positive supervisor and worker no-write events, a bound prelaunch log baseline, and exact worker/process/project/private-desktop identity. Potentially mutating supervisor actions and integration admission revoke this authority before dispatch. Proven no-write startup uses one 30-second cleanup budget and the existing five-second UI handoff; unproved ordinary cleanup and active-write waiting retain their existing policies. Ordinary global Stop still cancels its own queue.
- A late loaded ordinary target may enter one separate retained cleanup phase only after its original worker/helpers have exited and fresh complete loaded evidence is available. The existing observer persists the intent and absolute 30-second deadline in the existing session record before any action, under the existing native management mutex. Duplicate observers, explicit release and process restart cannot start it again. Prior No/Tip acknowledgements and the sole Close remain deduplicated. Unknown or multiple documents, changed identities and active writes authorize no action. Cleanup return never establishes release or successful verification; queued writes still use fresh scoped resource proof.
- Manifest hashing/copy/validation now run serially on the existing bounded daemon task. Nested executors no longer keep the Python interpreter alive after real MainWindow closure during stalled metadata I/O. Completed work renews inactivity; cancellation prevents late dispatch and incomplete copies remain available as evidence.
- Backup recovery no longer holds the global native mutex during abandonable metadata work. It retains the exact library lease and completed-work fences. Only the short pair of Work renames, rollback and installed-state update is protected from abandonment; the task registry guard is never held over filesystem operations. Cancellation after that pair cannot start target verification. An OS rename that itself never returns cannot be safely interrupted; the implementation preserves ownership rather than claiming cancellation completed.
## Implementation
- Resource release uses exact project/database/library scope and observed process identity. Read-only file probes retain exclusive handles through release commitment. Legacy receipts are rechecked; a proof for one block cannot close another block's session.
@@ -78,8 +87,28 @@ The final round-1 complete suite at `8818fa9` passed **2,358 tests in 451.34 sec
The missing frozen controls were exercised on B10: Open logs and Open exception evidence opened the exact current application folders, verified by private-desktop Explorer breadcrumbs; queued CBB recovery was cancelled before admission and created no recovery job; result Release opened the correct residual plan; Continue started only the second previously skipped item, which passed GREEN, without replaying the terminal first item. The first result summary elapsed under its existing three-second Close default. After ordinary main-window Release and normal dialog closure, production code recreated the Continue summary; the observer clicked that actual enabled control without injecting results or calling application methods. One bounded repeat captured result Release within the default window.
That button exercise exposed an **open ordinary-verification cancellation defect**: Stop 0.547 seconds after first PCB registration, before any Save, returned CANCELLED/safe=false at PCB_STARTING; the target later loaded Forward Annotation and remained alive. The current automatic loaded-startup cleanup requires diagnostic-only authority, so it did not clean this ordinary case. Exact manual Forward No and one normal Close after no-modal/no-progress checks were required before fresh release. The bounded result-button repeat reproduced the same retained state. These are failures of unattended ordinary cancellation, not successful automatic cleanup or protected active-Save cases, and remain for the whole-branch fix/review round.
That B10 button exercise exposed the **historical ordinary-verification cancellation defect**: Stop 0.547 seconds after first PCB registration, before any Save, returned CANCELLED/safe=false at PCB_STARTING; the target later loaded Forward Annotation and remained alive. The B10 automatic loaded-startup cleanup required diagnostic-only authority, so it did not clean this ordinary case. Exact manual Forward No and one normal Close after no-modal/no-progress checks were required before fresh release. The bounded result-button repeat reproduced the same retained state. These are failures of unattended ordinary cancellation, not successful automatic cleanup or protected active-Save cases, and remain preserved as failures; the grouped fix and B13 acceptance are recorded below.
The exact complete fault-bearing field snapshot is unavailable, so the original field inconsistency remains **NOT REPRODUCED**. Any bounded constructed live-copy experiment is reported separately. Windows 10 and production network/SMB behavior are **NOT RUN**. No merge, published release or issue closure is authorized by this candidate report.
One local B control required a second explicit verification after an initial completed native Save/Close left it PENDING. Both results are retained; automatic retries were not added and the initial failure is not counted as a pass.
## Final-wave evidence and limits
B14 EXE SHA-256: `0cd5f37b113802bcf3abd07a94d8c5b9f73d40696b1c56af66fedf39112e2d8c`. ZIP SHA-256: `a580e3e9945261761e0eb3968b4e271d9f4745ce982dea96c283af4ab06e565f`. All 68 packaged product modules/resources and all 242 extracted files match source/build. Five pre-cancelled frozen routes and common GUI/worker/owner-probe/dialog-agent smoke passed. Full-suite log SHA-256: `d9c57b7998e02154fb703f0fe2ba1f37f73dc4687d116c4b6efc0e5ad7e5c30e`; JUnit SHA-256: `4303cc123e90689103cc4135e483ac936b04a052067e952a1861ae167c7811ea`. Native clients and writer/management mutexes were idle before and after the full suite.
B13 (`5e9509e`) retains EXE `966e00fe57eccbdf4975c624026d39d24d4613f4174a932db248fb444a7caa9d` and ZIP `7a1da94e05b85873f791cde77907bdc280106768fbf84e5d89066ab8aa57a61b`. A final self-audit found that the new ordinary explicit-release branch bypassed existing cold-observer promotion. Its focused test failed before the two-line ordering correction at `c1e2596`; 110 related tests then passed in 21.96 seconds, including actual harmless worker exit, cold-observer promotion and queue proof. The in-progress B13 full suite was interrupted through its own execution session before editing, and is archived without a completion/JUnit claim. No additional native matrix was run for this narrow B14 delta.
The three review defects had retained RED evidence: actual MainWindow subprocess exit hung on stalled hash/copy/validation; real recovery callers held the unrelated native channel and could install after abandoned metadata; ordinary cancelled startup failed to clean a later loaded target. B13 related observer coverage passed 232 tests in 34.81 seconds, including an actual two-process mutex/file race between observer and explicit release plus a process restart. These race tests stub the native evidence/action boundary; they are not Mentor window tests.
B12 (`906cfc0`, EXE `7d7929d6621ae6c379205380e06f94eb076135343417d000018a8ab04efea8cc`, ZIP `bbe68005469d8b6cf87726200ef85646e442de45d164ab2ac0aec23b51d0bd65`) retained a new failure: ordinary `verify-642171759614` returned its visible cancellation terminal state in 7.797 seconds, but Forward loaded at 34.406 seconds, after the worker cleanup budget. Exact manual No/normal Close was used solely to clear that failed run. A later independent queue-helper attempt was refused with sharing violation 32 before native creation while two releasing observers raced for the coordination lease. Both failures are archived and are not relabelled as B13 passes.
B13 ordinary `verify-71aee0603183` reached the visible terminal state in 7.375 seconds. Its candidate worker acknowledged Tip at 21.359 seconds, Forward No at 27.422 seconds and one Close at 28.750 seconds; native exit followed the original cleanup deadline. Read-only sampling confirmed kernel exit. The original result remains CANCELLED/safe=false. No Release click or manual ordinary native action occurred. The helper stayed observe-only until the actual candidate produced fresh RELEASED proof, then obtained its own independent fresh proof. Its prequeued real Versions client `native-backup-b63e7f708fd5` was created 5.220 seconds after the candidate proof and 1.506 seconds after the independent proof, loaded the backup catalog and closed normally. This is **cross-owner queue acceptance**, using source-hash-pinned production queue/observer/native-worker modules; it is not survival of work in the same app after global Stop. The candidate itself performed every ordinary No/Close action.
The B13 native run issued Close within the first worker budget, so the new retained-phase No/Close branch was **NOT REPLAYED in Mentor**. That branch has source guard, real process race/restart and existing real Win32 adapter evidence. No repeated cold-start attempts were used to manufacture a slower launch.
B13 actual recovery `backup-restore-3f98efb14448` used the current pending original target and healthy independent source A. Its isolated source and installed target both returned PASS/GREEN/safe; native save/close/release completed without intervention. The old Work directory is retained and all 366 original source files are unchanged. The 27,689-file complete validation library copy progressed through the serial metadata path without a false inactivity handoff.
Separate frozen metadata-exit harness EXE SHA-256: `7490be1988d823944307510f7cfd22c85cf8570c7c25f6ee60d97f542bb72753`. All 68 packaged product modules equal final B14. A disposable I/O stall at each of hash/copy/validation exercised actual MainWindow close and interpreter exit before releasing the stall; no late dispatch occurred and partial copies were retained. This uses an acceptance-only entry, **not stall injection into the candidate EXE**. The earlier B12/B13 harnesses and all RED/setup attempts remain archived.
The original complete fault-bearing field snapshot remains unavailable (**NOT REPRODUCED**); the constructed partial copy remains **COPY_BLOCKED** and was never opened. Windows 10 and production SMB remain **NOT RUN**. The earlier first-pass B-control PENDING result and explicit retry remain separate outcomes. No merge, publication or issue closure is performed.