Hexen II post-notice startup and stdin ACK repair, 7 November 2026

The registered hexen2_demo passed its opening notice through ordinary Enter under cooperative scheduling, created a genuine native headless GL context, and opened MIDI with an observed function callback admission and subsequent return to inactive state. The paired run stopped on an UNKNOWN control reply; the real-thread arm was not launched. No reviewed menu or rendered world, auxiliary-thread execution, MIDI playback, gameplay, FPS or audio quality is qualified. This is an incomplete comparison, not a passing startup matrix.

Evidence: scratch/runs/20261107-graphics-post-notice/. Predecessor: notice-only sweep. The title's licence was respected: only runtime observations and emulator sources were examined; there was no guest disassembly, memory write, private input handler, forced fallback, guest setting/media/registry edit, clock override or nested worker.

Observed run and first failure

The original runtime source SHA was 94e52586d209dff22e1c3486fe3b45bf77be2b7f, with the previously published owning Worker MIDI repair present. Its relevant tracked source diff was empty; the existing untracked lib/wininit-boot.js is preserved as a new-file patch. runtime-pins.json, source-commit.txt and the source/input patches supply recovery through Git, without any source or media archive.

Local preflight checked the actual registry/media closure (16 paths, no missing or mismatched media) and literal relative CommonJS dependencies. Browserify virtual imports and one usage-comment literal were explicitly excluded from the filesystem audit. The production stdin client tests passed, including framing, exact IDs, errors, child exit and real Notepad input/capture. This preflight missed the interaction between trace windows and protocol output; the new regression below covers that demonstrated gap.

All 470 staged runtime/input pins matched before the guest and after its exit. All 148 observed runtime/native JS/JSON/WASM reads match their pins. The actual module was 2,182,493 bytes, SHA-256 e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467. The manifest, root executable and media hashes are recorded by existing path, without copying game media into local evidence. Root glh2.exe retains hash 76d96731e3076441e5c70bce2089cd8f36f25b373214f656aa9d5cd8d532e9dd; the registered layout avoids the bundled MiniGL.

Native preflight used Node v24.22.1, @node-3d/glfw 7.4.1 and @node-3d/webgl 6.0.3. A real context returned red RGBA [255,0,0,255], GL error zero and zero live contexts after destruction. The actual headless GL and WGL drawable/MakeCurrent/readback regressions passed. The provider was Mesa llvmpipe LLVM 20.1.2, WebGL 1.0, GLSL 4.50; 162 native dependency pins matched postflight. Native GL here uses Mesa software rasterization, distinct from the emulator's fallback renderer.

pair.cjs used the production test/run.js --control-stdin and shared test/control-session.js. Exact arguments are in retrieved/actual3/coop/launch.json: registered app, --no-build, --headless-gl, --no-threads, batch size 50,000, 150-second wall cap, normal CLI default clock, bounded API/input/scheduler traces through batch 15,000, individual PNG and PCM outputs. One guest ran at a time.

The 640×480 notice PNG was personally inspected: GLHexen2 accelerated-version information with OK focused. Ping, snapshot and notice capture received exact successful replies at batches 14,715–14,717. After review, Enter-down was submitted at 15:48:05.648 UTC and the independently armed Enter-up at 15:48:05.798 UTC. These are ordinary production control actions, with no guest command or private handler. Their ACKs were not observed. Release submission is established; final consumed/released key state is unknown.

At 15:48:06.131 UTC the title created a native 640×480 context with the provider above. WGL CreateContext and MakeCurrent both returned 1; no GL exception or fallback was observed. At 15:48:08.500 MIDI stream open returned handle 786433, CALLBACK_FUNCTION type 3, callback 0x441f70. The passive pump observation then records message 967 (MOM_OPEN), queue 1→0, callback-active 0→1 and EIP equal to that registered callback. The next observation finds the callback inactive at a different EIP. This supports open-callback execution/return; there is no enabled per-invocation return-token trace, streaming-header completion or playback proof. At exit the stream was not running, with zero pending headers/callbacks. PCM is an individually retained empty file.

The scheduler observer records backend cooperative, requested false, and zero auxiliary threads throughout. It provides no real-thread evidence. The individual raw guest log has no escaped trap or unimplemented API. No missing Win32/CPU semantic behavior was demonstrated.

At 15:48:12.655 the driver's unchanged 7-second reply deadline recorded UNKNOWN. It requested normal quit, saved the failure and skipped the threaded arm. No UNKNOWN input was replayed and no further original was launched. The final PNG was inspected: it is a 512×256 raw Hexen II logo texture selected from DX slot 3 by the harness, not a composited menu or world screenshot. The two PNGs were decoded and inspected; a texture does not establish visible menu progress. The planned post-notice control captures/snapshots are absent.

Demonstrated harness cause and repair

test/run.js emitted stdin replies with console.log. Its bounded trace window replaces console.log with a filter. Once this guest passed batch 15,000, successful and failed control replies were silently suppressed along with diagnostics. The absence of ACK therefore does not demonstrate a slow guest, emulator trap or unconsumed Enter. This corrects the initial timeout interpretation; the preserved UNKNOWN result remains authoritative for input custody.

The repair writes the existing [ctl] JSON protocol lines directly to stdout for both success and error. Diagnostic filtering, dispatch, IDs, input guards, deadline and scheduler behavior are unchanged. No API was stubbed or bypassed.

The extended test/test-control-stdin-cli.js runs actual production Notepad with closed --trace-from and --trace-to windows. It checks successful ping, synchronous-error replies, snapshots and ordinary Enter-down/up ACKs. Before the repair the first ping receives no reply and the self-bounded guest exits zero; the test fails with exited before replying to window-1. After repair both window cases and existing stdin tests pass. Shared-client and HTTP control regressions also pass. The mandatory flock /tmp/wine-build.lock bash tools/build.sh passed; the standard module hash remains unchanged. repair-source.patch records the tested repair separately from the original runtime's source state.

The remaining prerequisite is a fresh bounded paired post-notice run using this repaired stdin transport, preflighted with the exact closed-window arguments, retaining exact ordinary dismissal ACKs and composited screen captures. Native GL/MIDI open progress above is observed only in the aborted cooperative arm. This task stops at the first control boundary as requested; it does not repeat the accepted no-input startup matrix or infer parity.

Evidence and lifecycle

All 46 artifact references / 45 unique hashes were retrieved and size/hash verified before box stop, including the individual raw log, both PNGs, empty PCM, both WASM modules and native binaries. Runtime/native postflight mismatches and unmatched loaded reads are zero. Evidence schema/containment checks cover this run; unrelated global failures remain separately recorded.

A separate local lifecycle watchdog was armed before allocation with delegated normal-stop authority and a 25-minute cap. A persisted UUID Idempotency-Key created the only fresh box, bx_25tc5y5f, at 15:46:26 UTC, original expiry 16:11:26.724 UTC. No box was resumed. After retrieval, normal stop stop_837e40aeaa68b3cd completed at 15:49:38.286 UTC, forced:false, error null. The box is archived, stoppedAt 15:49:38.266 UTC, current archiveAfter null. No running runtime/box ownership remains to hand back.

Remote cleanup found no owned live process. No browser, profile or control port was created. Local lifecycle and messageboard watchers were closed and their PIDs are absent. Shared /home samples stayed above 2 GB; no peer files were removed. Only the harness repair, focused test and this note are published; the run-specific helpers and evidence remain in the ignored run directory.