Current Quake III browser startup, world and control boundary — 2026-11-08

The current module naturally reaches the native menu on both actual SoftwareGL and WebGL, and reaches the textured Q3DM1 world on SoftwareGL. The sweep's gray frame at 20 seconds was an incomplete startup observation. This run does not complete paired movement/fire/torch acceptance. The concrete WebGL blocker is the run observer's readiness predicate, not an observed emulator startup failure. Evidence is retained in scratch/runs/20261108-q3-current-browser-world/.

Inputs and provenance

The staged checkout is 44aca66a0c16ee9a002cb923228dcfbb7ca7cabd; the relevant tracked runtime diff is empty. relevant-source.patch, the untracked lib/wininit-boot.js new-file patch, owned-helpers.patch, helper pins and three runtime pin versions retain the exact inputs. Git remains the source archive. No emulator, acquisition, guest config, clock or tick code was edited. The guest's ordinary generated config is observed, never replaced or executed as an injected command.

The standard module is 2,182,500 bytes, SHA-256 f9650607181301b5d2a9990420f437bd9f1ca10d40ac73067339f07499d53121. The accepted Nov 7 firing/torch run used e415…; its gameplay is not current coverage. Both current source selectors are present in every original launch: gl-renderer and d3d-renderer. Actual contexts and draws determine the results below.

The complete existing registered quake3_demo closure was reused. There are 916 staged runtime/helper/media pins. The unchanged PK3 at test/binaries/candidates/quake-3-arena-demo-installer/installed/demoq3/pak0.pk3 is 46,853,694 bytes, SHA-256 e77abad2466f45a0a7ea018445528f9b95a0fe7789fa1abc1a7718bbf0754b08. The accepted 90-chunk, 524,288-byte transfer recipe verifies every chunk, assembles once, independently rereads the full file and removes temporary chunks. No source or media archive was created; media remains path/hash references.

Natural startup and actual renderers

One fresh UUID-keyed box runs the heavy browser sessions serially. The first software session bound its input identity to the initial GL endpoint too early. At 02:58:29 its reviewed frame is gray. By 02:59:20 raw diagnostics show a resident full PK3, no pending VFS read or read fault, changing guest EIP, 108,888 GL batches including 108,648 software-front batches, and presentation sequence 248. Q3 has created a later legacy endpoint for its software producer. The early binding cannot authorize input after that owner transition.

No guest input was sent in this first session. Its exact quit ACK precedes normal Chrome exit; controller/Chrome PIDs, ports and profile are verified gone before the second software session. The run observer is changed to wait for a live producer with triangles before binding. This is a harness correction, not an emulator fix or an original failure reproduced and repaired.

The second software session launches at 03:00:29.688 and its personally reviewed menu is captured at 03:00:59.478–59.826. A short raw observation at 03:00:51 shows PK3 seek positions, guest EIP 0x42cea6, software flag 1, a real software GL context, and advancing software-front batches. The initial native GL endpoint and the later legacy producer are retained rather than relabelled. FixedFunctionGL, context software=true, software-front execution, CPU triangle counts and advancing presentation sequence establish SoftwareGL; the D3DIM software-draw counter is zero and is not used as a GL draw counter.

Native menus retain Q3DM1: Arena Gate. All eight level-menu tabs toward Fight are individually acknowledged and reviewed. The initial Tab selected Multiplayer instead of Single Player; Escape and Up correct that ordinary menu selection. Additional difficulty focus corrections are retained; no console, map injection, private state or pause/freeze control is used.

After Fight, bounded ten-second passive commands show these distinct phases:

Capture time, UTC Personally reviewed checkpoint
03:05:07 Connecting to localhost, awaiting connection
03:05:43 Loading Q3DM1, maps/q3dm1.bsp, Arena Gate / Single Player
03:06:00 Loading cells
03:06:19 Textured courtyard, weapon, HUD ammo 100 and health 122

Raw diagnostics also retain changing PK3 seeks, guest registers/code/stack, no pending/faulted VFS read, GL triangle progression and zero fallback. This is natural load progression, not a frozen 20-second screenshot accepted as gameplay. The strict capture chain rereads actual CPU and render module bytes and maintains process/Worker/module/epoch/context/canvas continuity; both owning module hashes are the current f965… module in these software captures.

Actions that remain unproved

The acknowledged 1,200 ms MouseLeft action produces first-fire.png at 03:06:34 with ammo still 100 and health 66. Its camera differs from the first world frame. It is not accepted as firing or ammo consumption. A subsequent 1,800 ms W hold ends at 03:06:52 on “Fragged by Major”; that frame is not proof of forward movement. A separate 150 ms native click at the death screen does not produce a reviewed respawn. The first failed gameplay acceptance boundary is therefore the ammo-unchanged mouse action, followed by death; this run does not causally isolate guest mouse delivery or simulation polling. Turn and torch observations are skipped. No bot-damage or mouse implementation defect is inferred from these narrow observations.

An Enter respawn request at 03:07:32 is rejected before queueing because its 50-second ACK plus cleanup reserve cannot fit the immutable browser cap. deadline-refusal-proof.json proves no command file, ID file, transport failure or driver command failure exists for that request. The initial conservative local “stop” marker and its exact zero-effect correction are both retained. There is no replay. The 25 software commands have 25 exact ACKs; all admitted holds use independent release. Final raw state has empty input queue, all observed physical/async keys up, and mouse button mask zero. Chrome closes normally at its unchanged 03:08:30 cap.

WebGL observer blocker and skipped arm

After full software drain, WebGL launches at 03:09:02.530. The actual context is non-software WebGL 1.0 Chromium, supplied by ANGLE Vulkan SwiftShader Subzero. It is browser WebGL, not native llvmpipe or an Intel/Mesa claim. At 03:10:12 raw owning-renderer evidence shows 169,864 GL batches, zero software batches and presentation sequence 291, with no endpoint failure. The guest's generated config contains bind w "+forward" and bind MOUSE1 "+attack". The exact current standard module is recovered from instantiated Worker bytecode. Unlike the strict software capture chain, the WebGL observer never writes a completed per-owner module identity binding; the retained module file alone is not claimed as two independent owner proofs.

The personally reviewed composited menu at 03:11:17 shows Quake III's native Single Player/Multiplayer/Setup/Demos/Cinematics/Exit screen. Page owner fields are bracketed across capture. The guest and renderer are progressing, but the driver is still inside its run-specific producer readiness loop:

stats?.ready && stats.glTriangles > 0 &&
stats.renderEndpointId === a.renderEndpoint

For this WebGL GL endpoint, the page reports d3dStats: {}. The software-only counter predicate is false despite the real context, GL batches, presentation and visible menu. This is the precise causal boundary preventing the current driver from reaching chain.establishReady and its numbered input consumer. It is an observer error introduced by the software startup correction, not evidence that Q3's menu fails to load. No WebGL gameplay command is queued. WebGL Q3DM1 world, movement, turn, firing/ammo and torch are explicitly skipped. No second WebGL launch or replacement harness is attempted to disguise it.

Explicit Browser.close succeeds at 03:12:20; Chrome exits zero. The next passive readiness poll reports CDP UNKNOWN Runtime.evaluate at 03:12:41 after the intentional shutdown. This observer-after-close error is retained. Input is closed and no further backend or input follows. There is no UNKNOWN guest input ACK or unclassified escaped guest trap in the admitted stream.

Validation and lifecycle

Run-specific W/arrows/menu controls pass independent release under a deliberately slow sampler; UNKNOWN key and mouse controls prove their up events still run. JavaScript syntax checks pass. Terminal audit checks all three Chrome exits zero, all six process identities gone, six ports closed, profiles/writers absent and all 916 final remote pins unchanged. No emulator source changes are made, so no new build or failing-before/passing-after emulator regression is claimed.

All 28 retained PNGs pass chunk CRC and IDAT decoding. All 510 successful full HTTP reads match staged size/hash pins, and all 388 recovered artifact references (351 unique hashes) are independently reread before the stop request. The repository evidence checker retains seven unrelated containment errors; it is not reported as a clean global gate. This run's metadata references only its own retained artifacts.

The independently supervised 25-minute watchdog is armed before allocation. The sole box is bx_7hpx2hvn, allocated at 02:57:27.316 with initial expiry 03:22:27.174 and watchdog deadline 03:22:26.948. After independent retrieval verification, normal stop archives the box at 03:14:52.503. Final expiry is null. Receipt stop_bfb52f0a4e483c8f completes at 03:14:52.522, forced=false. Normal and watchdog delegates exit zero; the supervisor has no remaining children. Local free /home space remains above 5 GB. No RAM gate, resume, delete, source/media archive or public deployment is used.

Next work should make the existing readiness observer recognize active native GL ownership as well as the software producer, then verify original controls with trusted button delivery and actual simulation progress before claiming ammo, movement, turn or a torch view. This run establishes current software world and both menus; it establishes neither paired gameplay nor parity, FPS, audio, full playability or full27 completion.