Forsaken exit diagnosis and terminal-owner lifecycle, 2026-11-14

The original back-buffer error remains: ordinary EXIT GAME followed by promotion Enter reaches guest ExitProcess(13) at 214.580 seconds. No causal public-runtime violation was established, so no Win32, D3DIM or guest-code patch was made. The independently reproduced final-release lifecycle fault is repaired and validated against public process/HTTP fixtures and this original-game exit. OWNER_EXITED is explicitly neither physical-UP ACK nor a clean guest-Quit claim.

Evidence is scratch/runs/20261114-forsaken-clean-exit/; the final original is its validation/ subdirectory. causal-analysis.json, input-audit.json, lifecycle-summary.json and postflight.json contain the compact conclusions. The predecessor's UNKNOWN/child1, tool143 and missing additional waiter remain unchanged, hash-pinned in predecessor-preserved.json.

Original exit: evidence and remaining blocker

Bounded built-in API tracing covers calls 5175230–5175857, including the final consumed Enter, ordinary rendering, teardown and exit. Immediately before teardown, surface IsLost and presentation calls succeed. Teardown releases the D3D device/surfaces; RestoreDisplayMode5175737 and SetCooperativeLevel5175738(NULL, DDSCL_NORMAL) return DD_OK. Then:

Unmodified executable disassembly (exit-static-disassembly.txt) shows teardown 0x428ad7 invoking 0x4247d6, whose first store clears the game's render-ready flag at 0xe774b4 before external calls. Its render caller returns at 0x4274a1 and can continue to back-buffer show at 0x427501; the latter checks that flag before a public present call and reports failure through 0x427512. This is a static inference matched to actual return chains, not a live guest-memory observation or proof that native Windows follows the same earlier path.

PostQuitMessage queues a request and returns; it does not abort an active render call. PeekMessage retrieves WM_QUIT regardless of range. The known source range-filter discrepancy cannot explain this interval: preceding actual Peek calls use min0/max0, and there is no poll after PostQuit before failure. The SetCooperativeLevel contract permits NULL HWND with DDSCL_NORMAL; the RestoreDisplayMode contract does not establish the proposed synchronous callback needed to repair this continuation. No failing public present or demonstrated missing callback justifies changing those handlers.

The first causal public-runtime divergence therefore remains unestablished. An earlier callback/message differential or matching native-Windows comparison would be needed to distinguish the cause of this original continuation. Both authorized original sessions are spent; no speculative patch, forced exit0, third original or UNKNOWN restart was performed.

Lifecycle repair and tests

The supervisor now includes the actual launch PID start token in its native wait receipt, written after wait() and both output-drain joins. For a reviewed final action with consumed DOWN, the independent release child can record OWNER_EXITED only when that same PID/start identity has an actual successful harness wait, stdout/stderr EOF and no drain errors, within the existing two-second release cap. Endpoint closure alone is insufficient. Failed/signal waits, non-final actions, unresolved DOWN, missing EOF and mismatched identity remain UNKNOWN. The terminal records physicalUpACK=false, nativeQuitClaim=false and noReplay=true. Ordinary real UP ACK/progress behavior is retained.

Six new focused regressions pass: real normal final exit, TERM, KILL, non-final exit, late inherited stdout/stderr pipes with identity/invalid-receipt controls, and absent wait. Corrected red evidence reproduces UNKNOWN/child1 and missing wait identity before the repair. The first red invocation used a relative fixture-output directory and failed before DOWN; its infrastructure failure is preserved separately. The existing 29 controller tests and three production input-owner/guard/progress test files pass. Native headless GL, driver identity and the full heavy/build-locked flock /tmp/wine-build.lock bash tools/build.sh pass. Test logs and actual process waits are retained.

The additional run-local invoker catches TERM and preserves its actual child wait plus both EOFs. Real normal/TERM/KILL/late-pipe controls pass; KILL leaves the killed waiter's own receipt missing, without manufacturing evidence.

Bounded originals and evidence closure

The first fixed600/independent620 diagnostic started17:48:15 UTC. Bind was positively refused during startup, 3.186 seconds before actual native GL context creation: no owner, nonce or DOWN was assigned. The local controller stayed closed. Read-only capture and explicit host cancellation followed; orderly native harness0 is not guest Quit. The premature board assertion of successful binding was explicitly corrected.

The final fixed600/independent620 validation started17:52:10 UTC. It observed actual native GL context, successful submit and personally reviewed main menu before its single permanent bind. Nine manual400ms actions selected Normal VOLCANO, entered living play, paused, returned to main, selected EXIT GAME and dismissed promotion. Quit phase began with474.747 seconds reserved. Eight ordinary actions have real UP ACK, actual release-child0/bothEOF and subsequent progress; the final action has consumed DOWN then distinct OWNER_EXITED with actual release-child0/bothEOF. All captures were fresh within the supported 30-second menu window; all release/owner evidence was within2 seconds, with unique nonces, no replay and no UNKNOWN in the final run.

Final DOWN consumed at1794678945.634; release fired at1794678946.034095; actual native wait/bothEOF arrived at1794678946.060546 and OWNER_EXITED at 1794678946.072521. The failed release transport did not acknowledge physical UP. Guest exit13 is separate from native harness0. Both original chains have actual native/watchdog/supervisor/outer-child/additional-tool-child waits0 and separate EOFs; independent watchdogs did not fire and heavy locks were released. External execution-tool waits were actually reaped0 with combined EOF; their separate stdout/stderr EOF remains UNKNOWN because that interface merges output. All recorded native/outer PIDs are absent; closed port8527 is not ACK evidence.

All1648 pinned source files, both105-file loaded-source closures,894 media pins, 162 provider pins and tested module SHA256 c94ccf1b179f18bd4b47867445afed8cef2f1393b19aabfd259c7f3b94a7be09 match postflight. Sources are recoverable from exact Git SHA plus relevant binary patches, including the new test; no source archive, media copy or module bundle was created. Provider recovery used153 offline npm files with scripts disabled plus9 preserved native-cache files. The private installed modules were removed after verification; the original cache and root dependencies are preserved. Eight run-local inputs round-trip from Git patches. Ten named selected PNGs passed CRC/inflate/SHA verification; other captures were removed with hash records. Native logs retain every non-FS line with original line numbers.

The living scene shows lives5, SHLD128/HULL128 and PULS2000. This run sent no combat inputs. Prior outgoing-fire proof remains prior evidence; target damage, enemy kill, clean nativeQuit and full27 remain UNPROVED. ENEMY KILLED YOU is player death. No box was created: ID=null, state=NOT_CREATED, expiry=null, stop receipt=NOT_APPLICABLE. No deployment, workers or further originals.