Conan persistent preguest input release, 2026-11-17

The run-only Conan driver now initializes the proven persistent Q3 release service before launching the PE. Each gesture still obtains a fresh complete Conan D3D owner identity. The service source is byte-identical to the retained 21-gesture Q3 service. The first original subsequently exposed a separate production D3D frontend defect, repaired below.

The changed path completed 21 gestures on the public D3D9 fixture, with the real CPU Worker, render Worker, endpoint and ordinary compositor active. Presents advanced from 59 to 230. All 21 gestures completed independent physical UP, current-owner comparison, 500 ms prearm, 1750 ms UP acknowledgment and 3000 ms aggregate bounds. Strict native and whole-browser captures before and after those gestures passed. The complete after image was personally reviewed.

An unchanged real review aged 120004 ms and was refused before DOWN. During an intentional 350 ms producer stall, the independent service completed physical UP while the producer correctly refused its 200 ms DOWN bound. The initial fixture mistakenly expected that stall to be a positive; its raw terminal record remains intact and the expectation was corrected. A separate preparation was canceled before any guest launch.

Later disposable captures failed the unchanged three-second bound: a fresh capture after the age control, and a subsequent fixture's initial aggregate capture. Both fixtures closed immediately; no UNKNOWN was queried or replayed. The useful changed input proof combines the successful 21-gesture fixture, its real-age refusal and the earlier independent stall release. Further startup qualification loops were stopped. Existing complete D3D startup/capture and late/genuine-foreign-endpoint/closed/replaced-owner controls were rehashed and reused, as were the unchanged persistent service's cancellation/owner controls.

The served runtime remains exact b063d11ee8b53ca6ee4239f9c56b3f7691d4e9c2 with only da9ac26cc8c56a161b044e3705873420aecadaed host-imports, matching region map, and module 87ef39648a61ed6f54ca4ab8de1926de02c651eb2d99655015dc7e5de5407009. All 37 existing media/dependency files were rehashed without copies. Current foreign source and generated overlays were excluded from the served closure.

Evidence is retained in scratch/runs/20261117-conan-persistent-release-menu-gameplay/; final original outcome and native lifecycle receipts are recorded in its result.json.

The first original progressed naturally through Cauldron and NVIDIA. Four ordinary menu inputs completed physical UP using that same persistent service. Whole images personally confirmed the full main menu, New Game and Warrior selection. After selecting Warrior, guest thread 1 opened an instance-local GPS ring with resource ID 1, colliding with the live main-instance resource ID 1 in the shared render executor. Capture admission then became UNKNOWN and the run closed; no later input or query was sent. Movement, second gameplay response and native Quit were not demonstrated in that original.

The repair snapshots all 82 mutable GPS frontend globals under the existing reentrant process D3D lock, using the unused +40 slot in the existing 48-byte DX_PROCESS_STATE region. Each outer entry restores the process frontend and each outer exit saves it before unlocking. This shares ring sequence, resource IDs, targets, frame/coherency and caches across CPU instances without changing region declarations. The public front-sync export now uses the same lock. No scheduling, callbacks, input bounds or lifetime guards are relaxed.

The native two-instance regression reproduced the original duplicate-ID failure before the patch. With the repair it passes contention, retry, nesting, 4000 concurrent updates and 4003 real GPS CREATE/frame handoffs through the real stream consumer. Frame, lock/readback and render-target regressions pass against the exact affected module built from pinned b063 source plus only the two owned D3D source changes. Its SHA-256 is a441e4c89fa9912ddcc5778e1de90d5a795a2fb58d700104ab54de7184692085.

The repair is published on main as eca3d8f2f52aaede12fe716b39c88581a5e9c095, after the canonical locked build and the focused tests above passed. Strict pull and push succeeded. The commit contains only the two D3D sources, their existing regression test, the new generated layout record and the reviewed sync-front boundary entry. Concurrent COM review rows were preserved byte-exact through the short publication window.

The authorized affected original served exactly that a441e4c8… module, with base JavaScript plus da9ac, the unchanged base map and the original media. Its whole image showed the full main menu after 60 seconds of natural progress and no splash input. Enter and a pointer click on New Game both completed the same persistent service's physical UP/current-owner/deadline proof. Their subsequent images still showed the main menu, so no New Game response is claimed. A later strict native capture exceeded the unchanged three-second completion deadline. The run immediately closed UNKNOWN, with no later query, input or replay. This affected original did not reach Warrior or gameplay; movement, second gameplay response and native Quit remain unproved. The earlier original's reviewed Difficulty/Warrior images remain separate evidence, not an affected original acceptance.

A coordinator review identified a separate CLI host-export defect in the first repair: d3d9_gps_sync_front used the API lock's retry path, which changes EIP, yield and step state outside an API invocation. The actual-export regression confirmed those mutations. The corrective change uses a non-mutating host trylock and synchronous refusal on a foreign lock, preserving all 16 tested CPU/flags/control globals and 64 register bytes. Both ordinary and nested synchronous contexts pass; same-owner nested calls retain the outer lock and uncontended calls release their own lock. The native 4003-CREATE regression and the three exact-module frame/readback/RT tests pass again. The pure host path no longer reaches an EIP writer, so its now-obsolete producer addition is removed rather than granted ownership. The retained final corrective module is d42b335e08de55da44da04b8da8d84bf4ae9c76829d567d624727b1748fe78d5. The browser original did not call this CLI export and was not replayed after UNKNOWN. No browser claim is made for the corrective module.

Every browser/helper/driver/watchdog/supervisor/keeper was actually waited and both output pipes drained; retained native scopes and child sets are empty. Unique custody logs were retrieved and hash-verified before deleting only the owned tmpfs runtime caches. No box, worker, deployment, source archive or media copy was created. Final box ID and expiry are null; state is NOT_CREATED and box stop receipt is not applicable.

Next work is to resolve the affected original's unchanged-menu response and strict native capture completion under load, preserving all existing bounds. Only a newly justified, explicitly authorized original can extend the closed UNKNOWN evidence. Gameplay and native Quit remain the acceptance target.