Quake II original native GL encounter — 2026-11-14

Run: scratch/runs/20261114-quake2-native-gl-enemy/.

The original registered quake2_demo reached GAME → Easy, the opening room, the slatted passage, and the next tan industrial room. Living soldiers and incoming damage/death were reviewed. Three finite aimed Mouse1 bursts and one finite Enter attack pulse were delivered, but no distinct outgoing hit, defeat, or attack-triggered onward-door consequence is established. The native Quit confirmation completed with exit 0. This is a bounded gameplay blocker report, not full acceptance; Full27 remains unproved.

Exact route and recovery

The registry, media and bundled controls were unchanged. Startup was exactly +set vid_ref gl +menu_main, with authentic gamex86.dll, ref_soft.dll and ref_gl.dll preloads. No console/map command, forced guest call, authored state/config/AI/score write, tracking, autoplay, software fallback, extension, second game process, network download, box or deployment was used. The native game created its own entering save; after deaths, ordinary Enter and the native Load Game → ENTERING Outer Base selection restored it in the same process. No external save bytes were supplied or edited.

The retained Q3 native-recovery recipe supplied these exact manifest hashes:

Input SHA-256
package.json 3ed55cf48ef3a1a00e7971ea528b9b2c991decbbf8917901310d59f3808e936f
package-lock.json 125ea670ef6333b93d4375ed558fc42c24af2da9a4fbe2ed49e929ee519e0976

npm ci --offline --prefix scratch/runs/20261114-quake2-native-gl-enemy/native-deps --no-audit --no-fund returned 0. All 162 retained provider pins matched size and SHA-256 before and after execution. Actual runtime createRequire resolution from lib/headless-gl.js selected this run's private @node-3d/glfw and @node-3d/webgl. Root dependencies were untouched. The private node_modules was removed only after postflight verification; the package/lock, 162 pins and offline recovery recipe remain.

All 68 registered media/control/manifest inputs were present and pinned; no installer extraction or media recovery was needed. Principal hashes:

Input SHA-256
quake2.exe 2fa987c9390663e230c5b6479285e7c55511e35f80091bfed96f43062c1c810b
baseq2/gamex86.dll 11fd73be2887a6c567517dbbf74c2f5040c3636e4863befd1ade9f4537e52d60
ref_gl.dll d07b419749584220eb7748c781e181880a7919f38cf50e48b7c1e1f35772efcc
ref_soft.dll da4fd68c47a68712a8790aad3abb6a6b43355d0fc9d922fffea17998a421dbe7
baseq2/pak0.pak cae257182f34d3913f3d663e1d7cf865d668feda6af393d4ecf3e9e408b48d09
lib/quake2-modern-controls.ini afb166910b596b204e1c11f93ff4972999804d18299bde134d6380a51587e3b3

Under the heavy lock, all 317 accepted source pins and the built module matched the accepted SuperTux reuse chain, whose full locked build log is scratch/runs/20261114-supertux-original-consequence/build.log. No new full build was run. Actual loaded WASM SHA-256 was dcf6e8f0ca9d361c10fa556fb30835d34cf5bfe4089f84d298cb57fd170dbd42. Launch HEAD was 8c12efe999c1cf2a0079f427cf61416dee5ad129. Launch pins, relevant dirty Git hunks, new runtime input and new controller patches are retained; Git supplies source recovery, with no source archive. Postflight checked 317 accepted source files, 68 media inputs, 162 provider files and 103 observed loaded files with zero drift. Observed loaded files outside the accepted manifest are not represented as accepted build inputs.

The first native preflight passed: llvmpipe LLVM 20.1.2, WebGL 1.0. This run did not repeat the earlier Q3/Hexen II resize failure and does not fix it. Native execution recorded 4634 presents/readbacks and gl_sw_enabled=0. The llvmpipe provider is not a physical GPU or performance claim.

Reviewed gameplay and input limits

The one fresh runtime used --headless-gl --real-ticks, default 1000-block batches, cooperative scheduling, fixed --max-seconds=1200, and an independent 1220-second watchdog armed before launch. It took its fair heavy-lock turn after SuperTux enemy and TIM direction; the subsequent SuperTux worker queued behind it. Native launch was 13:45:50Z; stats report 1119.247 seconds and 961108 batches before native Quit. No deadline was extended.

The bundled ordinary bindings were inspected before launch: W/S forward/back, A/D strafe, Enter/Mouse1 attack, Space jump, Ctrl/C crouch. Keyboard turns and relative look were used. The native main menu was reviewed near shutdown, including the Options selection, but Customize Controls was not opened and no settings were changed. Bundled bindings alone do not prove guest action execution for every pulse.

Selected reviewed scenes include:

Navigation initially confused the near and far ends of the slatted passage; later scenes distinguish the tan room from the opening room. Some short pulses had little or no observed effect, and immediate captures sometimes preceded the final camera change. Later released/settled views were inspected, without promoting ACKs to movement or attack proof. Close melee and slow manual review left too little living time to establish a hit. No causal generic emulator defect was isolated, and foreign NE/render/host WIP was untouched.

The controller admitted 68 manually chosen scene-bound decisions, including 62 finite key/mouse pulses. Independent UP timers were armed before DOWN. Both capture-to-submission and review-to-submission admission limits were 30000 ms: observed maxima were 28093.304 ms and 605.165 ms. Capture time comes from completion of the actual PNG write, not a refreshed review. Each reviewed capture SHA and runtime PID/start token was bound to a one-use capture identity. A new review could not reuse it.

One later respawn request was positively rejected as INVALID, reason capture expired; NOT_SENT; it had no DOWN arm or delivery. A fresh full-scene capture and distinct intent then proceeded against the same runtime identity. The expired intent remains preserved. The command helper initially waited for nonexistent DOWN receipts on this refusal and timed out; the parent refusal receipt, rather than that timeout, establishes NOT_SENT. No UNKNOWN was cleared or replayed. Receipt totals: 211 ACK, 62 PAIR_ARMED, one INVALID.

The runtime used stdin inputs with controller PID/start-token and scene checks. It did not use guarded-bind/guarded-input for these gameplay commands; native queue/device/window epoch lease validation is therefore not proved by this run. Separate owner/guard regression passes do not establish that the live route used those guards. This is an additional qualification limit for any stronger acceptance claim.

Observer counts show 59 key DOWN deliveries, 58 key UP deliveries, and three mouse DOWN/UP pairs through takeInput; all 59 key UPs reached the renderer. These host input/guest-delivery observations are separate from Quake's internal attack consumption. No living attack pose or distinct outgoing hit is claimed. Final Y UP was acknowledged and reached the renderer, but the guest exited before consuming that last queued key-UP.

Tests, shutdown and missing requirements

Fresh/expired-capture-with-fresh-review, duplicate identity, new capture/new intent, wrong target, independent prearmed UP and terminal UNKNOWN controls passed. The expiry/identity cases passed before launch. A final prelaunch release fixture failed because its fixed 80 ms sleep observed the timer before its receipt-flush/finally completion. That failed output is preserved. The fixture was corrected to wait for completion and passed during the run; the runtime controller was not changed or relaunched. This ordering is explicit, not an all-prelaunch-pass claim. Three actual normal/deadline/signal lifecycle fixtures, current input-owner and guarded-input tests also passed. Native preflight passed on its first attempt. All 79 individually selected manual PNGs passed CRC, IDAT inflation and SHA-256 checks; they are not an automated frame series.

Native Escape → Quit → Enter showed QUIT? Y/N; Y completed native Exit 0. Actual runtime wait was 0 with separate stdout/stderr EOF. Watchdog actual wait was 0, reason runtime-terminal, with no deadline signals. The launch child and outer supervisor child also have actual wait 0 and separate EOFs. The watchdog inherited launch stdout/stderr, so its own separate EOFs were not captured independently. The detached outer process's own wait and auxiliary watcher wait remain UNKNOWN; their absence is not substituted for actual waits. All five recorded runtime/watchdog/supervisor/outer PIDs were absent, the heavy lock was released, and the watcher was signalled TERM. No box was created: ID null, state NOT_CREATED, expiry null, stop receipt N/A. Disk remained above the 1.5 GB requirement; no RAM gate or foreign cleanup was used.

Missing gameplay requirements are a distinct outgoing enemy hit/defeat or an attack-caused onward-door consequence, with living aimed execution and a reviewed released outcome. Stronger lifecycle/input qualification also needs actual live native owner-lease use and independent watchdog stream EOF capture. No full acceptance or repeat attempt is implied. The one bounded task stops with these requirements and evidence preserved.