Original Airstrike restored, visible combat and native Quit — 2026-11-14

The exact original closure is restored, and fresh unobscured combat and native menu Quit were observed. Production verification passes all 89 files / 4,072,418 bytes / 88 companion mappings, before execution and after both runs. The fixture is again available at test/binaries/candidates/airstrike-pre6a/. No emulator or registration code changed. Downloads and extracted media remain outside the run directory and outside Git.

Evidence: scratch/runs/20261114-airstrike-restore-combat/result.json, its selected PNGs, compact command/receipt JSON, and complete small runtime logs. All selected PNGs were inspected. The decisive fresh run is in fresh/.

Exact recovery

Only the two expressly authorized official URLs were downloaded, with HTTPS certificate checking, no redirect/provider fallback, bounded time and size:

Input Bytes SHA-256
https://icculus.org/airstrike/airstrike-pre6a-win32.zip 2,730,549 c611d1f269d2794920be7e715c40f251f93484b4ad00c399b4a9b8379aa2f662
https://www.libsdl.org/projects/SDL_image/release/SDL_image-1.2.3-win32.zip 160,219 e7f817604519192ee1be33f66190864898baf81fc87c36ff60e05f71e1901e14
Unchanged supplier zlib.dll 53,760 0df0c22bf411ad670d7b8fe5a533e4831313d5a3118cf9216389c642c90ec2f9
Standing test/binaries/dlls/msvcrt.dll 278,581 887eb5ce93edb7192ca3e9220f07f9ca0f94db02af5862ebcbdfcb852db99fd1

Archives, the selected original release tree and zlib are retained at test/binaries/media/airstrike-original-20261114/. Both ZIP CRC checks passed. Every selected original member was checked against the committed inventory before writing it. The inventory itself matches SHA-256 e7e62dd423c604effd0f747ea3f3c63ff445ee698b0755b52312b077ea481d75. The standing CRT matches the actual inventory pin; no provider substitution, rename, rebuild or media conversion was used. One acquisition-receipt script initially failed on a duplicate Python dictionary keyword after checking the CRT; the corrected receipt was written separately. Production preparation and verification independently passed.

The recovery recipe uses a fresh destination and these exact retained inputs:

node tools/verify-airstrike-pre6a.js --prepare \
  test/binaries/media/airstrike-original-20261114/release \
  test/binaries/media/airstrike-original-20261114/zlib.dll \
  test/binaries/dlls/msvcrt.dll \
  test/binaries/candidates/airstrike-pre6a
node tools/verify-airstrike-pre6a.js test/binaries/candidates/airstrike-pre6a

For reacquisition, retrieve only the exact URLs above into a fresh media directory, verify the archive hashes and CRCs, extract only members named by the inventory with source: "original" after stripping airstrike-pre6a/, and extract the supplier's unchanged zlib.dll. Verify every member's size and hash before preparation. Existing fixtures must not be overwritten. The committed verifier remains the authority for all bytes and mappings.

The release's own doc/airstrike.6 was read directly from the original ZIP; its controls section is retained as original-controls.txt: comma accelerates, Right/Left turn, period shoots, Space drops a bomb, and Escape opens the menu. Original airstrikerc, all data and sounds remain byte-identical: 800×600, windowed, alpha2, sound1, one human against the default AI. No source archive or additional supplier was acquired.

Fresh visible play and Quit

The decisive run used the production registered --app=airstrike_pre6a route, empty game arguments and original cwd, an existing pinned Wasm module, --no-close --real-ticks --no-threads --batch-size=20000, live ordinary input, and --screen=1024x768 before startup. Its wall-time guard was 160s and its independent observable watchdog was armed for 185s before launch. No guest memory, game config, clock, AI or score was edited. --no-threads uses cooperative guest scheduling; the original audio guest thread was still created and exited. Sound was enabled; audible output was not assessed.

The host initially opened the Console over the game. A titlebar drag moved it to x850, beyond the game's right edge x826. Clicking the visibly exposed game correlated focus and main HWND 0x10001. The fresh startup PNG shows the complete game area and “Press any key to start”; independently released Enter started the game. No host-window resize was used in this decisive run.

Selected full-desktop PNGs and actual settled batches:

PNG in fresh/ Batch Observed result
unobscured-start.png 35,668 Full original scene, startup text, Console aside
before-controls.png 38,238 Blue plane near (340,583), ground combat, score 00–00
firing-movement.png 39,207 Blue plane near (631,541), changed orientation/trajectory and white trail; black projectiles in front of it, explosions, score 01–00
post-release-combat.png 40,735 Continuing projectile/combat scene and blue plane near (600,571), after all gameplay key releases
native-menu.png 42,339 Entire actual menu, star beside Resume game, score 02–00 and combat wreck/effects
quit-selected.png 50,906 Entire actual menu, star beside Quit game

Comma was down at 38,920 and independently released at 39,928 (12s bound). Right was down/up at 39,025/39,079 (0.55s). Period was down/up at 39,091/39,747 (8s). Separate release processes were armed before each keydown; their parents actually waited for them and drained both output streams. The postrelease capture is 807 batches after the last gameplay release. These are visual observations of ordinary human controls during real combat, not merely input ACKs. Explosions, wrecks and score changes demonstrate combat consequences; no particular kill is attributed to the human's bullets.

After all gameplay keys were released, Escape down/up at 42,161/42,271 opened the freshly inspected seven-row menu. Its actual labels were Resume game, Players setup, Toggle fullscreen, About the game, Release notes, Toggle debug mode and Quit game. One Up down/up at 50,675/50,831 wrapped from the observed Resume row to the observed Quit row. Only after inspecting the star beside Quit did Enter down/up at 65,394/65,431 select it. There was no blind Down chain, control-channel quit or WM_CLOSE.

The complete log records [Exit] code=0, 65,649 batches and 94.137s. The runner actually exited0, both stdout/stderr reached EOF, its cooperative thread exited, and its process group was absent. The watchdog exited0 after the done marker; it did not fire. The heavy-runtime lock was released.

Diagnostic first run and validation limits

The first run omitted --screen; the host's initial 640×480 desktop caused the primary DirectDraw surface to be 640×480 while the game's native drawing surface was 800×600. Increasing the desktop after creation, attempted titlebar minimize/maximize and an ordinary host resize did not repair that existing primary. A malformed drag command was rejected without delivery, then explicit mouse down/move/up performed the titlebar drag. These diagnostic actions are retained, not replayed into the decisive run.

Read-only captures of the first run's actual 800×600 drawing surface exposed the clipped native menu and its selected Quit row. They are explicitly offscreen captures, not proof of an unobscured presented combat frame. That run also quit normally: [Exit] code=0, 257.346s before its 260s guard, runner0/bothEOF/watchdog0/no residual group. Setting the host desktop before the second launch resolved the presentation prerequisite without production changes. The second run supplies the full composited evidence above.

Both runs loaded Wasm SHA-256 c7f5bada4da4b4ed95b235add0aa1c9843176805d29d5b1ee79359752a19d348 (2,199,488 bytes). First-run inspected Git HEAD was 7a08e02858395b1b4c7e405fafef070d72cb4453; fresh-run HEAD was 85d1ddc92217807757bbebe718be01317940d6bf. Their source hashes and relevant uncommitted binary diffs, including untracked runtime-file patches, are retained in each run directory. Foreign WIP was preserved, neither edited nor adopted. These identify the tested existing module and runtime source; they do not claim a newly successful build from those dirty revisions.

Validation passed: real-input registration acceptance including all 89 missing-file controls, corrupt identities and mapping/seed/source/overwrite rejections; six existing menu-controller test groups; the independent deadline kill/wait/EOF control; the independent key-release transport control; and final production verification after both runs.

The required full-build attempt failed, under both wine-heavy-runtime.lock and wine-build.lock, at foreign TIM-owned ToyVM browser bundle freshness: 2,651,545 retained bytes versus 2,651,735 source bytes, first differences 801,005 and 853,049. build.log preserves the exact gate. No full-build PASS or production code commit is claimed. TIM subsequently announced its own regeneration/build success; that is not counted as this worker's test. This publication commits only this report.

All 89 fixture files verify unchanged after the final exit. No box was created: ID none, state NOT_CREATED, expiry null, stop receipt not applicable. Owned runtime/watchdog/release processes are gone; no runtime is retained. The assigned restoration/combat/native-Quit task is complete; no further gameplay or repair is pending here.