Cooperative GLTron sound startup, 2026-11-07

The unresolved boundary is GLTron's sound ring refill loop returning to its caller before the first scene draw. The SDL wave callback returns and its semaphore release reaches the auxiliary thread. No emulator repair is justified by this run. Cooperative menu and gameplay remain unreviewed.

Evidence: scratch/runs/20261107-gltron-sound-startup/ (ignored), following the presentation investigation. Only this note and owned run helpers were changed; production sources were inspected read-only. No nested workers, input, replay, configuration, clock, private-handler or forced-completion changes were used.

What the fresh observation establishes

One fresh UUID allocation, bx_cp7tgmst, ran the registered original through the predecessor's cooperative native-GL CLI command, adding only the existing --trace-callback-lifecycle diagnostic. The normal harness deadline was 125 seconds; the child exited zero after 125.655 seconds, with no pending command. All six compositor capture commands acknowledged. The seven personally reviewed 640×480 captures are teal at 5 seconds and uniformly black thereafter. No gameplay input was sent because no scene was available to review.

The passive preload preserves method receivers, arguments, calls and returns. It reads per-instance encoder getters, bounded guest stacks through the existing sparse-memory translator, the existing callback token ring, and shared audio queues. There are 126 actual-manager snapshots; utility-worker snapshots with empty managers are excluded. Sampling and diagnostics have overhead. Counts below describe this run, not equal-progress comparisons with another mode.

Boundary Observation
GL setup/publication 172 texture uploads, one clear, one guest swap/publication, zero submitted scene spans or vertices.
Per-instance encoder Main and auxiliary pending bytes/commands, spans and vertices remain zero in every snapshot. Final main counters: 342 calls, 341 flushes, zero spans/vertices.
Callback return 14,336 unique retained lifecycle events include 2,128 return events. Of these, 2,023 have their admission retained too; every pair restores saved EIP, ESP, yield and wait fields exactly. No sampled incomplete-return event.
Semaphore handoff 1,991 sampled grants and 1,991 grant-consumed events show auxiliary wait progress on 0xe0005. These are deduplicated sampled events, not whole-run totals.
Final callback state Active token zero, last WAT return token 136106, empty function-callback queue, two outstanding timed WAVEHDRs. A pending snapshot token at earlier times is followed by newer return tokens.
Actual ownership Main opens 22050 Hz, stereo, 16-bit output with function callback 0xa3e990; auxiliary writes/schedules 136108 buffers of 22048 bytes. Completion callbacks run on the registered main owner.

The trace ring retains only 128 events at a time. The sampled pairs prove those returns; they do not certify every callback or every interval between samples. The final token is a WAT return marker, not an inference from API counts.

Guest caller and first missing progress

The bundled sdl_sound.dll loads at 0x765000, original base 0x10000000; SDL loads at 0xa2e000. Read-only PE/import tables and retained objdump disassembly identify these boundaries:

All 112 sound-phase snapshots retain raw stack words 0x4132ac, 0x412c52 and 0x40cbf5; 106 also retain 0x766c02. Main EIP and saved callback continuations vary through decoder instructions between callback admissions. The stack scanner's candidate callers are heuristic; the raw stack and actual call-site disassembly are retained separately. In particular, its simple scan misses the three-byte indirect call preceding 0x766c02.

This places the missing observed exit in the sound refill/update path, rather than a callback stuck at its entry or buffered geometry awaiting publication. It does not establish that a single Sound_Decode invocation never returns: the same refill loop can call it repeatedly with the same sample and stack addresses. Decoder-call completion and ring read/write indices were not tracked individually. That is the precise remaining discriminator.

Completion semantics and timing

Microsoft documents waveOutWrite as submitting a prepared block and setting WHDR_DONE on completion; waveOutProc receives WOM_DONE when the driver finishes a block. These signals establish buffer ownership transfer, not completion of the caller's decoder or scene. The original SDL callback tests WOM_DONE and releases its semaphore. Its behavior was preserved, including its existing choice of synchronization API.

The inspected JS completion path marks DONE, clears INQUEUE and queues an immutable registration/header submission. WAT admission saves the interrupted register/wait context; the authenticated callback continuation restores it and records the return token. The scheduler grants the released unit to the parked auxiliary waiter, which then consumes its grant. The observed records agree with those paths; no wrong owner, lost sampled return, wait corruption or unconsumed sampled grant was demonstrated.

The unchanged CLI uses its synthetic batch clock (default 1000-block main budget, 200 ms per batch). At exit the audio clock reads 47,930,800 guest ms; this is not elapsed wall time. Timed completions remain queued ahead of that clock, and the function queue is empty at exit. The disparity makes producer / consumer pacing against decoder work a concrete hypothesis. It does not by itself prove starvation or an incorrect Windows primitive. No clock or budget was changed to make startup pass. A next investigation should passively pair decoder entry/return and ring-index changes with these existing completion tokens, distinguishing slow/incomplete decode from repeated refills that fail to reach the loop's exit condition.

Source, checks and handoff

Runtime source is main 1db931c53342c3bfc6b088fed1cecdf4fd238aa6, with exact runtime pins, relevant tracked diff and the existing untracked lib/wininit-boot.js patch retained. The actual standard module remains 2,182,500 bytes, SHA-256 f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121. All 129 observed runtime reads match their staged pins. Pre/postflight match 545 staged inputs and 162 native dependency pins. Existing passed controls were reused; headless GL and actual WGL context/current/readback checks pass. All 92 original-media paths are referenced with size/hash, without media or source archives. Node is v24.22.1, display :0, GLFW 7.4.1 / WebGL 6.0.3.

analyze.cjs passes byte/hash checks for every recovered artifact, callback restoration pairs, sampled grants, per-instance encoder state, all seven PNG pixel inventories, six acknowledgements and normal termination. Helper syntax and note whitespace checks pass. No production code changed, so there is no repair regression or new locked-build claim. This unregistered app's evidence stays with the investigation, not another app's dashboard entry. No historical census, parity, FPS or audio-quality result is claimed.

43 artifact references / 38 unique contents, 15,638,268 referenced bytes, were recovered and SHA-256/size verified, including complete compressed logs, stacks, captures and both modules. Git SHA plus relevant tracked/new/helper/note patches provide source recovery. Remote cleanup found no owned runtime processes, ports, profiles or held input. Launch free disk exceeded 5 GB locally and remotely (remote 57.27 GB).

The detached 25-minute watchdog and separate normal-stop delegate were armed and observed before allocation. Lifecycle ownership belongs to the coordinator: the inherited allocation helper's shutdownDelegated:true metadata is corrected by lifecycle-ownership.json; no normal stop was requested. After retrieval and cleanup the local delegates/watchers are retired for handoff, leaving the box running (API state idle), with original expiry 2026-11-07 22:51:06.810Z. No stop, resume, delete or public deployment occurred. Coordinator receives box ownership and the remaining expiry window. This bounded task ends with the sound-refill exit unresolved, not a menu/gameplay pass.

Coordinator lifecycle closure

After this handoff, the coordinator inspected the idle box and the43-reference retrieval verification, then requested normal stop stop_94b30b71284d081b. It completed2026-11-07T22:36:50.783Z, forced=false, error=null. Final inspected state is stopped, current expiry null. No resume/delete occurred. Exact CLI request/result and final info are retained in coordinator-stop-result.json and coordinator-stop-info.json under this run. Runtime findings remain unchanged.