Return Fire audio path, offline investigation, 2026-11-05

The concrete gap is an observation of successful audio initialization followed by a populated secondary-buffer Play and host delivery on the accepted route. No emulator defect is demonstrated; no production fix or audible-playback claim is justified. This assignment ran no game, browser, boat, capture or full build. Accepted launcher 5d014a0e and presentation/movement 9f695796c remain accepted.

Read first: CLAUDE.md, the Return Fire TODO, return-fire-demo.md, and alice-return-fire-reconciliation-20261018.md. Source inspection base: 58da3cdf95fb6c2f404e1d500fc8312efa6131d7. Relevant audio sources were clean; unrelated MDK/Stronghold work was left untouched. Evidence and executable offline fixture: scratch/runs/20261105-return-fire-audio-cause/.

The accepted October 31 run directories 20261031T2213Z-return-fire-presentation and 20261031T2230Z-return-fire-launcher are absent locally. Their complete dated notes were read, including original pins, ordinary audio/default Worker route, trusted input and normal cleanup; their screenshots/logs could not be reread. The historical 20261009T2052Z-return-fire-browser-audio directory is also absent. Its reported 99/100 non-silent analyser windows on module 527004d9... are historical, not current evidence or a listening review. Available worker completion JSONs only report exit status; they contain no guest audio trace.

Original input recovery uses existing paths, without copies:

All addresses below are original preferred VAs (image base 0x400000), not allocated emulator addresses. Use lib/pe.js VA translation or objdump -d -Mintel on the pinned original to recover the bounded disassembly saved in this run.

Stage Original call site / prerequisite Current implementation
Startup 401358 and 402754 call 4240d0(hwnd) Native startup path, no title override
Timer / format allocation 424118: SetTimer(hwnd,1,100,NULL); zero result takes failure path. 424132 calls thunk 427d08, IAT 47b468: acmMetrics(NULL,50,&size); nonzero fails src/09a3-handlers-audio.wat: $acm_metrics writes size50, returns0 for metric50
Device 4241ad -> thunk 427d02, IAT 47b33c: DirectSoundCreate(NULL,&[463440],NULL) src/09a8-handlers-directx.wat: $handle_DirectSoundCreate
Primary 4241c7: SetCooperativeLevel(hwnd,2); 4241fd: CreateSoundBuffer, legacy descriptor size20, flags1, NULL format, result at 45fd50 Existing cooperative-level and CreateSoundBuffer handlers, including legacy20 descriptors
Failure Any initialization failure reaches 424208: warning at 472390, then releases device/primary/format and kills timer Guest explicitly permits continuing with no sound; a visible game is insufficient proof
Load effects 4209c0 immediately skips if device 463440 is NULL. Table starts 467bb8, stride28, 37 entries. 420a29 calls 421660, which calls 421420 RIFF/fmt reader and 4215b0 buffered data reader WINMM mmioOpenA/Descend/Read/Ascend/Seek/GetInfo/Advance/SetInfo/Close in 09a3-handlers-audio.wat
Populate 420a94: CreateSoundBuffer(size20,flags0xe0,data length,loaded format); 420acb: Lock; guest copies PCM; 420b0b: Unlock; 420b26/420b41: volume/pan0 Existing DirectSound buffer format, Lock/Unlock, gain/pan handlers
Request playback 420840 returns early if byte 4633e0 is0 or device 463440 NULL. 4208bc: DuplicateSoundBuffer; 420904: secondary-buffer Play (vtable+0x30), reserved args0 and flags derived from caller $handle_IDirectSound_DuplicateSoundBuffer, $handle_IDirectSoundBuffer_Play
Host Play obtains voice and calls host_voice_play_ring with backing bytes, start offset and loop bit lib/guest-rpc.js brokers initial Play to page; lib/host-audio.js voice_play_ring -> _voices.playRing -> PCM decode/BufferSource or eligible looping worklet -> voice gain/pan -> wave bus -> master -> destination

Startup's separately allocated default format is PCM stereo22050Hz/16-bit (bytes/sec88200, block align4). This is not the effect-file format. The 37 original RFIREDAT/SOUND/*.~AD files are RIFF/WAVE PCM tag1, mono8-bit; 35 use11025Hz, JEEPSTAR uses5563Hz and RELOAD uses11127Hz. Each has a complete data chunk containing bytes differing from unsigned PCM silence128. For example HELI has data offset58, length17407, SHA256 0ccfb24afbe30c796f83d38a99990f919f05c167a7fe579e09664f6860ac2859. No ADPCM decoder or missing effect assets are demonstrated. A string Music/RFireScore.aiff exists but that asset is absent from the original inventory; a string alone does not establish an executed open or explain silent effects. MCI imports/call sites are separate from the demonstrated DirectSound effect path.

The small offline fixture reads originals in place, checks all payload hashes, case-normalized VFS coverage, RIFF chunk bounds, effect formats/data, all37 table references, and nine exact instruction pins. Its first run incorrectly assumed all effects used11025Hz; that failed assertion is retained. The corrected fixture validates each file's own byte rate/block alignment and preserves unusual rates. It establishes the actual inputs/checkpoints for the missing observation, not execution of guest calls. No permanent regression test is warranted without a demonstrated implementation failure.

Validation: node scratch/runs/20261105-return-fire-audio-cause/offline-fixture.cjs passes (58 payloads,37 effects,37 table entries,9 instruction pins). node test/test-audio-worklet-routing.js passes36/36; node test/test-audio-mixer.js passes its four groups. These are isolated mocked host checks, with no shared module mutation. Existing DirectSound format, volume/pan, replay, loop-refresh and mmio tests were inspected as relevant coverage; WAT compilation suites were not run. No code edit/full build was needed or attempted before the Stronghold final runtime/file RELEASE gate.

Next bounded observation, for a separately authorized runtime after lane release:

  1. Preserve the unchanged original58-file manifest and pin the actual served source/module/Worker route. Use the accepted ordinary Play/H/W inputs with defaultViewport:null, indexed8 and default audio. Retain API entry and return values from startup through the first H effect, with a hard30-second audio observation window and normal cleanup. Do not redo launcher/paint fixes.
  2. Record SetTimer result, acmMetrics50 output/result, DirectSoundCreate result and 463440, primary creation, each effect open/load byte count and secondary creation/Lock/Unlock result. At the effect request record 4633e0, device, DuplicateSoundBuffer/Play HRESULT, flags, format, byte length and PCM hash. Stop at the first failed boundary. No calls means inspect the guest gates; a failed open means inspect VFS requests before blaming host routing.
  3. If Play succeeds with populated bytes, observe the page's matching voice_play_ring, source scheduling, AudioContext state/currentTime progression, wave/master gain and mute, and existing _wineWaveAnalyser / _wineMasterAnalyser signal. Trusted pointer/keyboard gestures drive _voices._installUnlock; a suspended context must be distinguished from missing requests. Gain-aware mixer peaks can reflect queued input PCM, so they alone do not establish destination delivery. Capture/listen only under separate authorization; even a running nonzero master analyser does not prove audible output, quality, looping correctness or synchronization.

Evidence stores Git SHA plus relevant binary diff/new-note patch and input path/hash inventory, without source/media/module snapshots or archives. Runtime, box ID/state/expiry: N/A. Coordinator retains lifecycle/shutdown ownership.