Forsaken original native Quit, 2026-11-14

Evidence is in scratch/runs/20261114-forsaken-native-exit-repair/. This investigation uses the unchanged original demo, synchronous native D3DIM, real ticks and the ordinary SinglePlayer / Normal / VOLCANO route. It makes no white/blend, game-state, input-owner or dependency changes.

The fixed 600-second diagnostic reproduces the stuck Quit. Its bounded API trace shows PostQuitMessage(0) (API 5804323, return address 0x428b28) returning before the original caller constructs the back-buffer error and calls MessageBoxA (API 5804333, return address 0x428bc3). The earlier block watch observes bRenderingIsOK change from 1 to 0 during shutdown; a block watch does not identify the exact writing instruction. Static original disassembly shows shutdown at 0x428ad7 calling renderer cleanup, then posting quit, and the render caller falling through to ShowBackBuffer after shutdown. The modal is parked at yield15 with explicit quit pending. causal-exit-extract.log retains the actual ordering and original x86 frames.

This distinguishes an original post-shutdown rendering attempt from the emulator's modal-loop defect. Microsoft's MessageBox explanation says MessageBox exits its modal loop for a pending WM_QUIT. Modal dialog guidance requires preserving quit for the outer loop. PostQuitMessage queues that request; a pending flag alone does not prove native process Quit. The existing shared modal pump never checks this pending explicit request.

The diagnostic also shows living traversal toward the hatch and a nearby green enemy craft. A finite aimed primary Ctrl2000 pulse had an independently armed release. Its actual midpress capture shows death/respawn and PULS2000; the postrelease capture shows four lives and PULS1949. Ammunition use is proved; projectile impact, enemy damage and a kill are unproved.

A local action name pause-up collided with the previous action pause's UP receipt filename while saving a postrelease capture. The action had already been consumed, independently released with ACK, and made postrelease progress. A read-only production nonce query confirmed consumed/CANCELLED with reason independent release, never UNKNOWN. Original failure and resolution records are preserved; no action was replayed, owner rebound or guard changed. The run-local helper now rejects output-name collisions before any intent or DOWN.

Repair and public tests

The uncommitted repair in src/09c3a-dialog-runtime.wat ends the owning modal wait only for the explicit PostQuitMessage marker 2, using the existing modal teardown and saved-call continuation. It leaves quit pending for the outer caller. It does not manufacture an OK button, successful render result or process exit. The speculative window-recreation marker 1 remains unchanged. The only new runtime branch is eight lines in the modal pump.

test/test-modal-postquit.js runs synthetic x86 through the real public PostQuitMessage, MessageBoxA/W and PeekMessageA thunks. The final test is red on the original pump (done=0, quit=2, modal=65543, yield15) and green on the repair (done=1, quit=2, modal=0). It checks the real stdcall stack and outer WM_QUIT retrieval. Ordinary dialogs with markers 0 and 1 still wait for Enter. Eight adjacent suites also pass: native modal command routing, A/W MessageBox, Worker shadow input, Worker common dialogs, EndDialog lifecycle, teardown quit, main-window recreation and Win16 modal destruction. repair-checks.json records nine successful actual waits with separate stdout/stderr EOFs. The public documentation predicts the Windows modal behavior; no original-game comparison on a Windows installation was performed.

Concrete publication blocker

No repair code is committed or published. The required full locked repair build returned 1 at the callback ownership census before compilation. It identified the changed modal pump and the separate LMarbles worker's changed renderer producer. The census working file already contains six unrelated foreign review/hash entries for uncommitted thread, render and NE sources. Committing the whole census path would adopt those unrelated entries; publishing the modal source without its matching review would break the clean-main build gate. The task requires explicit-path commits and preservation of foreign WIP.

Only the modal pump's own census row was updated and reviewed, then released uncommitted. census-own-review-preservation.json proves the existing foreign hunks were preserved at that update. The LMarbles owner subsequently updated its own renderer row separately. The exact own source/test patch is owned-code.patch, and the exact modal census row is owned-census-review.patch. The repair, test and review row remain coherent WIP in the shared checkout; they must be published together using a method that preserves the unrelated census changes. No build gate was disabled.

The repair build is not claimed successful, and no 900/920 validation was launched. The next step is to resolve the owned review row's publication scope, pass the full heavy/build-locked build, publish the modal repair with that row, then perform the still-unused bounded original validation. Native Quit after the repair remains unproved.

Original outcome and retained evidence

The original diagnostic's launch build passed under the heavy and build locks in 61.564 seconds. The loaded module SHA256 is dcf6e8f0ca9d361c10fa556fb30835d34cf5bfe4089f84d298cb57fd170dbd42. source-pins.json, actual-loaded-inputs.json, relevant Git binary patches, 894 media/input pins and 162 private offline provider pins identify the actual inputs. Postflight immediately after the diagnostic found no source, loaded source, media, provider or module drift. Subsequent repair/foreign edits are separately recorded; they were not retroactively attributed to that run.

The original finished on max-seconds600, with runtime stats599.996 seconds. Native exit0, the independent prearmed620 watchdog's unfired exit0, supervisor, outer-child and invoking-tool waits were actually observed with their output EOFs. Timeout and exit0 are not native Quit. The input audit contains twelve unique consumed actions, one permanent owner, independently armed UP/ACK and postrelease progress (the final reviewed choice instead awaits actual process completion). Maximum scene age was12.429 seconds and review age0.534 seconds. The primary pulse lasted2.000095 seconds; its capture completed inside the press, before independent release.

A test-only fixture initially overlapped another worker on port8466; that fixture evidence is invalid and retained as a failure. Final controller tests use exclusive8486 and verify the actual fixture child PID. Eight controller, two extension, four supervision and one outer-wait controls pass. Production input used direct8477 throughout and had no UNKNOWN receipt or owner change.

result.json names the eight retained full PNGs. Their dimensions, CRCs, inflation and SHA256 are verified in png-retention.json. Repetitive API output was compacted after its EOF; the original byte count/hash and selected line ranges are retained. No source tree, media, module binary, frame series or capture archive is retained. The private provider is reproducible from its small pinned offline recovery inputs and was removed after use. No box was created: ID/expiry null, state NOT_CREATED, stop receipt not applicable. Full27, target damage and native Quit remain unproved.