Quake III Arena demo 1.11

Display gamma presentation (2026-10-19)

OpenGL frames now opt into display gamma in the presentation filter. It reads the WAT-owned SetDeviceGammaRamp tables through dib.gammaLut for the composited window image, so GL attachments/readback and canonical GDI pixels stay intact. Both local GL and render-Worker frame layers carry the OpenGL marker; other graphics APIs keep their existing gamma handling. Identity ramps keep the direct-canvas bypass. The new-player seed retains r_gamma 1.3 but drops the temporary r_ignorehwgamma 1 workaround; persisted configs are not rewritten.

The pixel regression test/test-opengl-gamma-web.js exercises actual local WebGL and a real render Worker frame, distinct RGB tables, identity restoration, repeated presentation, untouched source/GL readback, every scaler with and without an exclusive viewport, dedither, both correction backends, and exclusion of DirectDraw and hidden windows. Q3 q3dm1 acceptance uses hardware gamma explicitly enabled and paired captures with only the display correction toggled. Evidence: scratch/runs/20261019-q3-display-gamma-no-threads and scratch/runs/20261019-q3-display-gamma-threads-verified. Those captures predate the 2026-11-05 restructuring below and were local Chrome/SwiftShader checks, not Safari.

2026-11-05: the first version was an extra pass over the finished, scaled presentation canvas (texture upload of the whole output, draw, copy back). Headless Chrome on an M1 (ANGLE Metal, alternating arms, loaded box) measured it at +0.7 to +1.0 ms/frame at 1x and +1.8 to +3.4 ms/frame at 1920x1080 (throughput). The tables are now applied where each path already handles the source: GPU scalers (FSR1, auto pixel scale, dedither) correct the source texture in a source-sized framebuffer pass; Canvas2D scalers draw through an SVG feComponentTransfer table filter that is used only after a 256-entry ramp drawn through it matches the tables exactly, otherwise through one WebGL copy of the cropped source at source resolution. Measured delta now: +0.02 to +0.04 ms/frame (filter) and +0.05 to +0.2 ms/frame (FSR1) for 640x480 and 1024x768 sources at 1x and 1080p; the WebGL fallback costs +0.65 to +1.5 ms/frame, almost all of it the WebGL to Canvas2D hand-off. Smooth scalers now interpolate corrected source pixels rather than correcting interpolated ones.

Worker joining and download deadlines (2026-10-16)

The remaining c29ca48c1 discovery/join changes are integrated while keeping the newer shared Q2/Q3 account-name policy. A nonempty hosting directory is offered at launch; empty directories stay silent. Native discovery joins adopt the room on the existing guest and send net_restart; connect, including punctuation, rather than restarting Q3. Reviewed local isolated-Worker evidence: scratch/runs/20261016-q3-worker-join, using the deployed WASM with the candidate JS changes. Host .1, discovery client .2, and unchanged client guest identity are asserted; this is Chrome-to-Chrome coverage.

Lazy HTTP file/part downloads now have a 90-second total deadline covering headers and the complete body. Timeout aborts the fetch and rejects even if an adapter ignores abort, so existing parked-read retries and Retry/Quit UI can run. Late responses cannot populate the cache. This deadline is for provider-backed game data; eager downloads and bundle requests are separate.

Brighter first-launch defaults (2026-10-15)

The original pending config seeded r_gamma 1.3 with r_ignorehwgamma 1 so the engine applies gamma during texture uploads. Gamma alone had no visible effect on the deployed runtime's display-ramp path. Saved player configs still replace the seed. No texture-intensity override is added. Before/after Chrome captures of q3dm1 are under scratch/runs/20261015-quake-brightness/quake3_demo-{before,after}-verified. The comparison uses the stable deployed WASM with intercepted seed configs.

File-access trace (2026-10-15)

Reviewed scratch/runs/20261015-file-reads-quake3_demo/: q3dm1 rendered on the stable deployed WASM/local range server. pak0.pk3 is 46,853,694 bytes; successful guest ReadFile calls cover 11,348,898 unique bytes. The current cache fetches 109,768,254 bytes in 105 requests, with repeated chunk loads. Guest reads span all five 10 MiB deployment parts. Retaining each whole part once would cost 46,853,694 bytes. This is one first-map trace, not a runtime speed benchmark; the hosting card remains open in the final screenshot.

Hosting banner follow-up, 2026-10-15

The shell previously equated owning a browser LAN room with running a game server. Q3 opens its socket at the main menu, so it showed the hosting/share card too early. Probed games now wait for a positive hosting probe; the card is offered once per room. Q3_HOSTING_CARD=1 node test/test-quake3-desktop-web.js verified no card at the menu, a card after starting a server, and no reopening after dismissal across probe refreshes. Reviewed captures: scratch/runs/20261015-quake3-host-card.

The reported client stuck at “Awaiting challenge” happened with the host already in a map, and Q2/Blobby failed too. An isolated-worker probe confirmed the shared defect: browser/assigned IP was 10.0.0.2, while the guest main worker retained the default 10.0.0.1. Initialization and joinVlan updated the idle browser instance and child threads but skipped the main worker. Both paths now propagate the address; startup LAN pacing also reaches the worker. WORKER_VLAN_ONLY=1 node test/test-worker-guest.js covers startup assignment, late reassignment, and pacing on an actual isolated worker.

Reviewed scratch/runs/20261015-quake3-worker-join confirms host .1, client .2, both executing in workers, open WebRTC datagrams, and server-side ClientBegin: 0 plus ClientBegin: 1. This focused run verifies joining and gameplay rendering, not mobile input. The original desktop harness now accepts Q3_WORKERS=1, serves isolation headers, and asserts the execution mode and actual guest IPs. Earlier runs without isolation exercised cooperative mode even when no-threads was omitted. Safari-to-Chrome was not tested locally.

Cross-game checks with --isolate: Q2 room test passed all 35 checks, including four players, late console join, and invitation-link join. Blobby passed 32 of 33: discovery, both players' movement, slow-peer traffic and rehosting passed; hidden-host rendering produced 2/4 distinct sampled frames instead of the required 3/4. Its peer kept moving and traffic continued. That remaining background-rendering check is not marked passing.

App id quake3_demo (2026-10-10, w4). OpenGL only.

Provenance and license boundary

Desktop preparation, 2026-10-14

The first-launch lib/quake3-modern-controls.ini is mounted as c:\\demoq3\\q3config.cfg; saved user configuration replaces that seed on later launches. W/S move, A/D strafe, Space or right mouse jumps, Ctrl/C crouches, left mouse fires, wheel changes weapons, E/Enter uses the held item, and Shift walks. relativeMouse: true uses the same browser capture policy as Quake II. Mobile has a held move pad (Up/Down also navigate menus), relative look surface, and Fire/Jump/Crouch/Item/Weapon/Esc buttons.

The browser LAN uses an automatic room on UDP 27960. Its hosting probe sends getinfo berrry and reads the real server's hostname response; the room card launches clients with +connect {host}. No engine patches or native UDP Internet access are implied: the transport is the existing browser WebRTC virtual LAN. The standalone CLI pair remains test-quake3-vlan-gameplay.js.

test-quake3-desktop-web.js exercises trusted browser keyboard/mouse input, measures movement and turning with the game's own viewpos, then joins a separate mobile browser through the normal room card and checks the server's ClientBegin log, bidirectional WebRTC traffic, touch movement and look. Reviewed acceptance: scratch/runs/20261015-quake3-acceptance6/ (UTC date; 2026-10-14 locally). Chrome with SwiftShader, cooperative mode, desktop plus touch-emulated mobile browser contexts:

Canonical build, asset/manifest-size checks, FPS touch-layout and file-policy tests pass. The deploy dry run considers 2262 files; all 303 installed manifest entries exist, and pak0.pk3 is emitted as five parts. No public deployment was performed. Physical-phone/Safari behavior and WAN/NAT conditions were not tested by this local two-browser run.

Running-game joins and native server browser, 2026-10-15

The release acceptance originally covered the pre-launch room card only. Q3 also opens a socket immediately after an offline launch, producing another room picker after command-line arguments have already been consumed. The running-game join recipe now handles that socket-time picker (why=1) and the later invitation (why=0); the pre-launch card still uses +connect.

Original demo Con_ToggleConsole_f at 0x406640 toggles bit 0 of cls.keyCatchers at 0x6dc1a4. Renderer startup creates the window before the UI is ready, so the socket-time recipe waits for a nonzero catcher before opening the console. Q3 consumes printable text through WM_CHAR, unlike Q2's key-event recipe: type(text, { characters: true }) explicitly opts in to character delivery. No guest memory is patched.

Reviewed scratch/runs/20261015-quake3-running-join4: offline launch, socket-time room join, console connect without launch arguments, two-player gameplay, native localservers discovery, mobile movement/look/fire and portrait controls all passed. Earlier runs exposed the second picker, premature key injection, and missing character delivery.

Reviewed scratch/runs/20261015-quake3-native-browser: joined the browser LAN with automatic connect disabled, opened the original Multiplayer menu, selected Local and clicked Refresh. The actual native list displayed NONAME, Q3DM1, 1/4 DM, UDP 155, and 1 of 1 Arena Servers. The native localservers handler at 0x40ee70 broadcasts on ports 27960–27963; cls.numlocalservers at 0x6dc250 became 1. The Internet master-server list is not the Berrry room directory: native discovery sees peers in the browser LAN room, not every published Berrry room.

Probe-gated room creation (2026-10-15)

The demo's socket routine at 0x4415c0 maps its localhost bind string to INADDR_ANY (0.0.0.0), not loopback. An open UDP socket therefore cannot distinguish the menu, solo play, and hosting. SVC_Info at 0x433ab0 reads g_gametype and returns without replying when it equals 2 (single player). The constant at 0x4b137c is float 2.0. This makes the normal getinfo probe a useful hosting signal without inspecting guest memory in production.

Q3 now enables lan.openOnProbe: it initially receives a local-only wire, and a valid infoResponse promotes it to a public owner room. Native Local discovery can open the room picker; choosing a room restarts the guest with the assigned member IP and existing connect arguments. Explicit invite links continue to join directly.

Reviewed run 20261015-quake3-probe-room3 passed in isolated Chrome workers: menu and spmap q3dm1 (g_gametype=2) kept no public room, hosting card, room URL, or networked background policy. A multiplayer map promoted to owner .1; a separate menu stayed offline despite that host, then native localservers opened the picker and joined gameplay as .2. A third browser joined gameplay through the invite as .3. Safari cross-browser behavior was not tested in this run. test/test-vlan-star.js also covers response validation and discovery filtering without public peers.

Q3 sends Local discovery automatically on entering Multiplayer, before the player chooses Create or Join. Empty directory results therefore stay silent. When other hosts exist, the picker also offers Host a game, which closes it and leaves Q3's native Create button available without opening a room. The dismiss action is Back to game, since it does not disable networking.

Install

The package is Installer VISE (MindVision); neither unzip nor 7z opens it, so the original installer runs in the emulator: node test/run.js --exe=.../Q3ADemo.exe --vfs-include=**/* --batch-size=10000 --control-stdin --frozen. Welcome -> Next (452,393); Software License Agreement (scroll the text by clicking the scrollbar track at (568,250); Page Down only moves the caret) -> Yes (452,393); destination c:\Q3Ademo -> Next; Ready To Install -> Next; copying demoq3\pak0.pk3 (46,853,694 bytes) takes ~50k batches at --batch-size=10000 (about two minutes); Finished. The 304-file c:\q3ademo tree was exported from ctx.vfs with a control eval (keys are lowercase) into test/binaries/candidates/quake-3-arena-demo-installer/installed/, and node tools/gen-installed-manifest.js <installed> quake3.exe wrote the 303-entry .wine-assembly-browser.json. No registry state is needed.

Running

Whole-part HTTP validation (2026-10-15)

scratch/runs/20261015-whole-part-games/quake3_demo reaches q3dm1 with the whole-GET provider. All five simulated 10 MiB pak0.pk3 deployment parts are requested once each; no HTTP Range requests. Stable deployed WASM, cooperative Chrome/SwiftShader, not a timing or multiplayer benchmark.

Current native browser arena (2026-11-12)

Current native arena checkpoint records the original registered demo reaching Q3DM1 through Single Player, Fight and Difficulty on both software and WebGL with actual Workers and the current standard module. Separately reviewed world images show balanced W movement and native left-mouse firing (ammo 100→98 software, 100→96 WebGL). For this unchanged browser seed, MOUSE1 is attack and Ctrl is crouch; the older direct-map/CLI control observation above is historical. Both browser and runtime parents exited zero with stdout/stderr EOF. No production defect or runtime fix was demonstrated. FPS, visual parity, audio and full corpus coverage remain open. Evidence: scratch/runs/20261112-quake3-current-native-arena-0204/.