Recoil demo

Source: https://archive.org/details/RecoilDemo, original RECOIL.zip. Archive metadata declares 30,351,332 bytes and SHA-1 ac0f23a00dc9e132d7abd21ab07934b0ef0adbc4. No title was present in apps, the candidate manifest, tasks or the board when the source lane was claimed on 2026-10-11.

The ZIP central directory contains the HTML instruction manual in DOCS/ and recoil.exe (29,250,305 uncompressed bytes). The extraction below is verified; gameplay is not. Keep the title local-only unless its distribution terms are established. Do not answer installation approval dialogs automatically.

Controls from the included manual

These are documented bindings, not observed emulator input responses. Any eventual touch layout should use the actual bindings and be tested against player-controlled gameplay.

Preflight

Remote box bx_y3exnm39, source aa7838d8c, full build passed. The archive download endpoint returned HTTP 500; direct mirror whole-file and 4-MiB range downloads timed out. Resumable 256-KiB ranges completed and the full archive matches the declared SHA-1. The WinZip wrapper contains an InstallShield5 cabinet; tools/is5-extract.js extracts 107 files, 76,007,851 bytes, including RecoilDemo.exe, native MFC/CRT DLLs and zbd/ data. Local test/binaries/candidates/recoil-demo is absent by design because the persistent box has less than 2 GiB free. Native fallback files test/binaries/dlls/ole32.dll, test/binaries/dlls/shell32.dll, and test/binaries/tlbs/stdole2.tlb are absent on the tested remote box.

Current startup result

The original executable SHA-256 is 82632b1b46153014af0caf1e0c9dfdebbb0d8482ba65b2735064f10972afa076. Tested WASM SHA-256 is 7c5968c791e2310e16fd14bf1b4eca43e4ff758b35c55ad354b0f895666a3361.

CLI startup renders the splash and a graphics selector. Ordinary clicks at (168,289) select software, then (323,326) accepts that graphics setting. Game at (25,31), Start at (38,52) reaches "Failed to initalize video device". The accelerated choice reaches the same message. Neither is gameplay. Runs: 20261011T1853Z-recoil-startup, 1854Z-recoil-software, 1855Z-recoil-menu, 1856Z-recoil-start-game, 1859Z-recoil-hardware. Short run names retain the same date prefix.

The API trace narrows the software route: DirectDraw2 SetDisplayMode requests 640x480x16; primary CreateSurface succeeds (EAX=0 at return0x49be20, output0x08140048). Surface3 QueryInterface succeeds with output0x08219020. The original surface is released; the game then tears down the new interface and restores the display, without another traced surface setup call. Do not mislabel this as CreateSurface failure or change its successful return. Next identify the post-QueryInterface/Release failure condition using the existing API/return tracing. No renderer fix or root cause is established.

Candidate preparation and the existing touch-controls and proprietary dropdown tests pass. The touch layout is based on the manual, not a tested mobile play session. Registration remains local-only; FPS/audio unverified.

Browser launch comparison

Run 20261011T1903Z-recoil-browser-dlls uses Chrome 152.0.7977.42, Threads and the lazy candidate manifest. Explicit preload of the bundled MFC42/MSVCRT/MSVCP50/MSVCIRT fixes the earlier generic-CRT lookup traps (run 20261011T1901Z-recoil-browser). The reviewed screenshot still shows "This program requires the file COMCTL32.DLL, which was not found on this system." This is earlier than the CLI graphics selector/video error; subsequent scripted clicks did not reach a menu and do not qualify as input or gameplay evidence.

Source inspection explains the lookup difference: browser-shell.js advertises all dll-registry.js DLL_PATHS as available, including COMCTL32, without testing file presence. host-imports.js has_dll_file therefore returns true; the worker LoadLibrary path fetches the missing native image and returns zero. The CLI instead reaches the existing static-system COMCTL32 handle 0x5d110002. The tested boat lacks test/binaries/dlls/comctl32.dll; the bundled Recoil archive does not supply it. Next distinguish incomplete fixture provisioning from desired fallback semantics with a matched browser/CLI fixture. Do not add a silent success stub or claim the CLI video error is fixed.

Final registration checks: remote test-touch-controls.js and test-local-proprietary-demo-dropdown.js both pass at 19:08:49Z.

Video-init cause and surface-interface fix

The COMCTL32 browser failure was test-box provisioning: restoring the native DLL alone still404s without the normal binaries -> test/binaries symlink. With both restored, reviewed run20261011T1914Z-recoil-native-alias reaches the same video error as CLI. COMCTL32 is no longer missing on bx_y3exnm39.

The guest helper at0x49be10 creates a surface, queries Surface3, then releases the original interface without resetting EAX. Its caller0x49c0ee treats any nonzero result as failure. The breakpoint at0x49be45 proves EAX=1, not a CreateSurface or QueryInterface failure (run20261011T1916Z-recoil-release-return).

Surface3 now uses the existing DX auxiliary-interface reference accounting. AddRef/Release return that interface's count, while the total still controls teardown, implicit attachments, fences, clipper references and pixel storage. The original interface can reach zero while Surface3 retains the live surface. This follows the separate ref1/ref3 behavior in Wine's reference implementation: https://github.com/wine-mirror/wine/blob/master/dlls/ddraw/surface.c (see ddraw_surface1_Release and ddraw_surface3_Release).

Extended complex-surface regression tests the QI/AddRef/Release sequence and final storage refund. Old code fails (3 instead of2); fixed passes. Final scoped patch full build and complex-release, Surface3 descriptor, clipper, and surface-fence regressions pass on the boat. An initial broader prototype reaches the rendered New Game menu (run20261011T1918Z-recoil-surface-fixed); that capture is not gameplay and predates the final Surface3-only scope. Final-patch New Game validation continues separately.

Final patch validation: run20261011T1920Z-recoil-newgame (CLI) and run20261011T1921Z-recoil-browser-fixed (Chrome152, Threads) both render the main menu without Video Error. Run20261011T1922Z-recoil-browser-newgame holds a click at the visible New Game center (page380,340) for150ms; WM_LBUTTONDOWN/UP reach HWND10001 but the menu does not advance. Next inspect fullscreen client-coordinate conversion and the game polling path. No gameplay/FPS/audio/mobile qualification.

Relative mouse menu route (2026-10-11)

Recoil registration/refcount fix integrated on main ed013f819. Browser run1927Z confirms delivered pointer358,106 and HWND10001; held click/Enter do not advance. Main-menu polling calls DirectInput GetDeviceState16 bytes at return0x46690c (buffer0x564230), not an absolute Win32 cursor. Relative movement and held button work: run1931Z reaches name/difficulty. Read-only cursor data at0x564258/0x56425c shows (357,109) before a move and (460,479) after (82,313); game sensitivity scales motion and clamps. Button byte0x56423c is0x80 while held. No input emulator change was needed.

Verified CLI sequence (batch-size default1000; source9c7cbf823): 120:click:168:289,130:click:323:326,160:click:25:31,170:click:38:52, 20000:relmousemove:30:-134:10,24000:relmousemove:0:36:10, 26000:di-mousedown:1,27000:di-mouseup:1, 62000:relmousemove:66:255:10,64000:di-mousedown:1,65000:di-mouseup:1. At180000 batches reviewed run20261011T1934Z-recoil-level-route shows the 3D opening: tank on a truck beside the water. This is not yet gameplay. Longer650000-batch run finished at19:36:14Z with no reported crash; its 450000/520000 captures and held-Up460000..510000 require separate review. Browser ordinary gameplay, FPS, audio and mobile remain unverified.

Player control verified in CLI run20261011T1938Z-recoil-turn-check: the 450000-batch baseline exactly matches the forward-run before PNG; holding Left460000..500000 visibly rotates tank/view left and moves the obstacle to the right edge, with minimap orientation changing. Reviewed before/after gameplay captures are published. Browser/mobile/FPS/audio remain unverified.

Browser relative input (2026-10-11)

The registry opts into relativeMouse because Recoil menus and aiming consume DirectInput relative motion. The existing browser pointer-lock path supplies unbounded mouse deltas. No renderer or emulator core change is needed.

Reviewed run20261011T1954Z-recoil-pointer-files on Chrome152 with Threads and cross-origin isolation reaches New Game, name/difficulty, mission screen, then gameplay HUD. Ordinary Puppeteer mouse motion and held clicks drive those menus; read-only guest cursor feedback guides the test. No guest state is patched. The initial Left hold overlaps the cutscene transition, so this run alone does not establish browser player steering.

The fork omitted prepared candidate files and native COMCTL32. Restore real files under test/binaries/candidates/recoil-demo/game; the emulator server rejects symlinks outside the checkout. Regenerate the browser manifest with the existing candidate preparation tool. The earlier files-unavailable runs1947Z/1951Z/1953Z are fixture failures, not pointer-lock regressions.

Existing touch-controls and local-proprietary-demo-dropdown tests pass. The touch test needed its original packages/freeware/blobby-volley/settings.dat fixture restored on the fork. Mobile behavior, FPS and audibility are unknown.

Browser player control verified in run20261011T1956Z-recoil-browser-turn. Wait100 seconds after Start so the opening finishes, then hold ArrowLeft for2.5 seconds: the tank/view turns left, the wreck shifts to the right edge and the minimap rotates. Both frames show gameplay HUD and enemies. Chrome152.0.7977.42, Threads and cross-origin isolation enabled; WASM a2eca9be09f0f966f019e61fa02fd0aafef61008517c155a33e78de9a8495eeb. This qualifies browser steering, not mobile controls, FPS or audibility.

Phone route follow-up

Touch-enabled Chrome at844x390 renders the pad and Fire/Alt fire/Map/Menu controls (reviewed run20261011T2007Z-recoil-phone-start). Switching a live desktop session into touch emulation did not instantiate the controls; run2004Z is a harness limitation, not evidence of a broken phone launch.

Touch-only run20261011T2009Z-recoil-phone-menu remains in the native graphics selector: a200ms touch at307,214 selects software, but the same held tap at421,240 on OK does not close it. Later menu taps consequently test nothing. Capture filenames describe intended checkpoints; reviewed contents show the selector still open. Next trace native-dialog touch/button dispatch, including hit coordinates and down/up pairing. No root cause yet; phone gameplay/steering/FPS/audio remain unverified. The desktop route is unaffected.

Native control touch fix

Run2012Z traces the failed OK touch to the old radio cursor266,288 and HWND10005. A physical mouse at the same screen point reaches420,323 and HWND10006, closing the selector. Applying existing main05cac1070 (Drakan modal-touch guard) still fails this selector in run2015Z.

The touch bridge now checks its existing native-control hit predicate before choosing trackpad/aim-only behavior. A native button receives the finger position and a held down/up pair. Game surfaces retain relative input. Reviewed run20261011T2017Z-recoil-native-control closes the graphics selector with the same touch. Regression fails old source at20,30 versus405,225, passes fixed for relative and aim-only profiles, and preserves non-control tap semantics. Edge-release and web-touch-input checks pass. The separate web-touch-cursor script skipped because its Chrome detector found none; the Puppeteer real-page replay above ran Chrome152 successfully.

This fixes the phone selector; touch-only gameplay is still pending.

Native menu touch follow-up

Run20261011T2022Z-recoil-touch-route exposed the next native UI target: Game-menu taps reused cursor419,323 after the selector closed. The native press predicate now recognizes menu bars using WAT-owned geometry and open menus (including taps that dismiss them), before native child controls. Regression fails old/passes fixed, covers scaled coordinates, absent bars, open menus and outside-presentation rejection. Touch-input and single-app layout checks pass.

Reviewed run20261011T2024Z-recoil-touch-menu-fixed reaches the game main menu through touch-only Software/OK/Game/Start. Tapping New Game does not advance: that menu belongs to the game and polls its DirectInput cursor. Next verify ordinary touch-trackpad drags and tap at the resulting cursor, then gameplay. No phone gameplay/FPS/audio claim.

Touch-only gameplay verified (Chrome emulation)

Reviewed run20261011T2033Z-recoil-touch-held at844x390 with Threads uses only CDP touch events: native Software/OK/Game/Start, trackpad drags to New Game and Start, and800ms held Fire presses on the overlay. Cursor coordinates are read for test feedback; no guest writes or keyboard/mouse injection. After100 seconds, a2.5-second Left hold on the direction pad rotates the tank/view and minimap, moving the wreck to the right side. Before/after gameplay captures are reviewed.

UX follow-up: ordinary trackpad taps reach and highlight New Game but do not activate it (run20261011T2029Z-recoil-touch-gameplay); held Fire does. The remaining issue is the short trackpad click pulse, not cursor movement. This route verifies Chrome touch emulation, not WebKit or a physical phone. FPS and audibility remain unmeasured.

Short trackpad tap fixed

Recoil opts into touchControls.trackpadTapHoldMs=100. A trackpad tap now keeps mouse-down for100ms so frame-polled input can observe it; other apps retain their existing immediate tap. Requested holds are capped at500ms. Pending release is paired once and cleared on blur, hidden document, or new touch/mouse input. Native direct controls retain their existing path.

Reviewed run20261011T2043Z-recoil-tap-hold opens the name/difficulty screen with an ordinary trackpad tap, where baseline2029Z only highlighted New Game. Regression fails the previous source and passes fixed, including deferred release and idempotent blur cleanup. Full touch gameplay was verified in 2033Z using held Fire; it was not repeated with this timing patch. The boat expired after evidence collection. Next validate the full ordinary-tap route and measure FPS/audio; physical phone/WebKit remain unverified.

Current-main ordinary-tap gameplay qualification

Run 20261011T2056Z-recoil-main-tap-gameplay rebuilt and tested main 274902b0258a26c41ac03b6bb190e1fbbc7c0c05 on temporary boat bx_k7xdfrq9. WASM SHA-256: d5199d347404b8562a67dcf60e0e18ad5896e1d6a2bcd4c4de1a42559676b5b9. Chrome 152, Threads enabled, touch viewport 844×390: Software → OK → Game → Start, then trackpad drags and ordinary taps on New Game and Start reach gameplay. The held-Fire workaround is no longer needed. A 2.5-second Left pad hold turns the tank/view/minimap; reviewed before/after captures show the wreck moving from ahead to the right edge. No guest writes or keyboard/mouse injection were used.

During the following 30.002255-second gameplay sample, the existing WinePerf.presentsTotal counter advanced by 453: 15.10 guest presentation events/s, with 453 canvas uploads. HUD drawing and input tracing were off. This is not a qualified logical-frame or physical-display FPS measurement; Chrome ran headless with SwiftShader and the game selected its software renderer. A source download for the next title overlapped on the same boat, so this observation is not a controlled performance benchmark. Raw counters, build log, route, source identity and reviewed screenshots are in the run. Audio, WebKit and a physical phone remain unverified.