Quake III pinned browser startup, 2026-11-17
Run: scratch/runs/20261117-quake3-pinned-browser-startup-arena-backends/result.json.
This is generation e7c13db296e380130bc6b9771cf3f55970f5c42f, module
f40fc54c5b0f1d4bd505364049e091c0092cab27eac0dcb644e8623d2344875b;
it is not a latest-main runtime claim. The prepared published Git source,
module, registered quake3_demo caller, and 307 input/media paths were checked
offline. Media was referenced in place, without copying or downloading it.
The preceding sweep attempted 24 logical cells across six titles, not 27 applications. Its threaded Q3 gray/black endpoints and cooperative intro textures did not establish a CPU crash or an active owning executor. Those results and their interrupted/missing waits remain unchanged.
One fresh registered cooperative original was used for each requested backend. Natural startup continued beyond 20 seconds under a persistent local supervisor and independent deadline/disk-floor watchdog. Ten-second passive observations retained EIP, yield, guest mode, actual bridge/executor identity, GPS counters, and browser/Worker events. The original module was also recovered from live CDP bytecode and matched the prepared hash.
Both cells reached the full main menu. The WebGL cell first reported an
actual GpuStreamWebGL.Executor on its gl-gps bridge at about 50.8 seconds;
the software cell reported an actual GpuStreamSoftware.Executor. Increasing
GPS record/present counts and reviewed composited menu images accompany those
identities. Neither a renderer string nor the mere existence of a page WebGL
context is the executor proof.
WebGL stopped when its own watchdog measured 1,499,545,600 free bytes, below the owner's actual 1,500,000,000-byte floor. One reviewed menu pointer click had occurred; no arena or native Quit was reached. The later CDP session-not- found error followed cancellation and is not a guest CPU-crash diagnosis.
After space recovered, the unused software original ran. Reviewed ordinary
inputs advanced through Choose Level and native Game Server setup with Q3DM1
selected, Free For All, and Back/Next controls. These images are menus, not
arena gameplay. A later Tab key dispatch exceeded the harness's independent
1,000 ms hold interval. The harness recorded UNKNOWN and closed the run.
Its finally release path is source evidence; that failed action has no
retained successful per-action release receipt or proven guest consequence.
There was no replay or extra original. This establishes a harness/input
observation limitation, not a demonstrated production CPU/renderer defect.
A separate initial browser observer error occurred before any guest module
or game-media read (runningApps was not yet defined). Its logs and actual
close/drain receipts are retained separately; it was not a guest execution.
Both original Chrome processes exited normally with both pipe EOFs. Driver, watchdog, and persistent-helper actual waits and both EOFs were retained; owned browser scopes were empty and disposable profiles removed. The detached root custodians have no parent wait receipt: their absence is recorded without substituting it for such a wait. No box was created.
No production correction was justified, so there was no canonical rebuild or post-fix original. Arena movement/turn/fire/enemy consequence, native Quit, gameplay parity, FPS, audio, threaded behavior, and the wider 27 remain open. Both original backend authorizations are now spent. Further work should first qualify ordinary input delivery/observation under the actual software page load, and obtain separate authorization before another original execution.