ScummVM — Flight of the Amazon Queen (floppy)

Binary: test/binaries/candidates/scummvm-fotaq/scummvm.exe (SDL 1.2.9 build, DX5 video driver) + SDL.dll, queen.tbl, games/FOTAQ_Floppy/queen.1. Registered in lib/apps.js as the local candidate scummvm_fotaq (same files and command line as below; the candidate binaries live only in the main checkout, so a fresh worktree needs them copied or symlinked into test/binaries/candidates/scummvm-fotaq/).

node test/run.js --app=scummvm_fotaq \
  --quiet-api --no-close --batch-size=25000 --max-seconds=580 --input=...
# equivalent, without the registry:
node test/run.js --exe=test/binaries/candidates/scummvm-fotaq/scummvm.exe \
  --vfs-include='SDL.dll,queen.tbl,games/**' --args='-pC:\games\FOTAQ_Floppy queen' ...

Module layout

Fixed (2026-09-29)

With these the intro, cutscenes, autosave, credits and the first gameplay room (hotel room, Joe, verb panel, inventory) are reached at about batch 8700 with no crash or unimplemented API.

Fixed: mouse input starved in the room (headless)

pollEvent ran about twice per 200 batches (once in 11,100 batches from boot), so posted WM_MOUSEMOVE never became an SDL MOUSEMOTION that ScummVM reads. The earlier theory — that $host_get_ticks stepping the headless clock kept a timer always due — was refuted. The real cause was MsgWaitForMultipleObjects answering "message waiting" forever: both its handler and the $win32_dispatch fast path tested $nc_flags_count, which stays nonzero while any window holds an erase bit (2) or the persistent default-erase bit (8), neither of which PeekMessage ever returns. SDL 1.2's DX5_CheckInput (sdl+0x10015090) keeps pumping while MsgWait says yes, so DX5_PumpEvents never returned. Fixed in 1279f024 ($msgwait_queue_ready, counting only NC paint/calcsize bits via nc_flags_scan(5), the same test $has_pending_message uses; covered by test/test-nc-flags-message-wake.js). After it: 1,836 pollEvents in 3,000 batches, the intro and credits play, the hotel room is up by batch 8,000, hovering the chest shows "look at chest", and mousedown, gap, mouseup walks Joe to it.

Browser audio stall on the Worker backend (2026-10-03)

(Recovered from branch archive cbfe9895, Oct 2026.) In the browser Worker backend, submitted PCM stalled at 44,096 bytes (~0.5 s at 22050 Hz stereo16): SDL.dll's waveOut CALLBACK_FUNCTION at original VA 0x10012140 releases the audio semaphore on WOM_DONE, and Worker execution then had no completion-queue delivery (the cooperative loop pumped it), so both completions sat queued and the audio thread starved. The required shape is delivery on the owning Worker, never fire_wave_out_callback on the page shadow instance. (Likely fixed since: lib/guest-worker.js now calls fire_wave_out_callback_bound.) Also observed: the registered command line has no -f, and Alt+Enter did not toggle fullscreen in that check (undiagnosed).

GOG SDL2 build

The GOG ScummVM (SDL2, the one that actually imports GetMessageExtraInfo) dies earlier on InitializeSListHead (in a DllMain) and then VerSetConditionMask. Not addressed.

2026-10-10: CLI walk and a qualified present counter

The registered SDL1 app at 25,000-block batches, with Escape about every 1000 batches, reaches the hotel room by batch 9027; a held floor click at (420,290) (down 9037, up 9043) walks Joe from x178 to x555, and an --input replay reproduces the frame byte-identically. Evidence: scratch/runs/20261010T0640Z-fotaq-control-frames.

Over batches 9027-9400 all 4,098 slot-6 presents are IDirectDrawSurface_Blt to the primary (0x080e2010) from one SDL.dll site (return 0x9851db, sdl+0x100181db): one per ScummVM updateScreen. ScummVM pushes the screen every engine loop, so only 243 of them (5.9%) change the picture — read the present count as the engine update rate and the changed count as animation.