Quake III fullscreen/output visibility investigation, 2026-11-15

No general production defect was reproduced. No production correction or fresh original was attempted. The missing evidence is the preceding original's final failed fullscreen/output/occlusion operands. Its three-second screenshot UNKNOWN and compositor cause also remain unresolved.

Evidence: scratch/runs/20261115-quake3-fullscreen-visibility/result.json. The controller and native fixture diagnostics below are tested additions, not a repair or evidence of menu, arena, frag or native Quit progress.

Prior original remains closed

The complete retained image was personally reviewed: stone-texture startup, host consent strip and desktop/debug chrome; no Q3 menu or arena. host-fullscreen-observations.jsonl records a fully visible output immediately after trusted host consent. Subsequent startup observations report ordinary damage and held/in-flight paint, then failure.json reports startup owner/output not live and fully visible. The final failed visibility operands were not retained. The earlier body-level touch-cursor overlap cannot identify that later cause: the existing guard already excludes ordinary elements outside the actual fullscreen subtree. The production cursor is appended to document.body, outside #screen-wrap.

Failure screenshot 1465 became permanent UNKNOWN after three seconds and produced no recovered pixels. Neither original route was replayed. The measured native shader status queries are legitimate work and warrant no rendering, synchronization or deadline change.

Added diagnostics

tools/quake3-visibility-controller.cjs composes the existing controller using exact, fail-closed source anchors. Every passive page observation includes selected output and fullscreen rectangles, backing dimensions, ancestor geometry/styles and client bounds, intersecting light-DOM boxes and their ancestors, pseudo-element metadata, top-layer elements, five hit-test stacks, viewport/focus/pointer-lock state and all 256 physical key states. The original owner receipt remains attached. A Runtime binding emits the operands before the unchanged readiness predicate can throw. Missing bindings fail closed.

Hit testing supplements conservative painted-box detection because an element with pointer-events:none can still occlude pixels. No cursor exception is added. The diagnostic adds no canvas-pixel, guest-memory or GL-state reads and changes no DOM, production styles, rendering options, input state, capture predicates, protocol deadlines or independent release behavior. Metadata serialization and extra DOM/style reads have a scheduling cost; zero perturbation is not claimed. The only executable mode is synthetic native qualification; there is no original launch mode in the run recipe.

Actual qualification

Heavy work began only after explicit StarCraft mission213262 and Sims brush242527 source/build/runtime releases. Their receipts were personally verified: 325 StarCraft waits/PIDs and 29 referenced Sims wait files/87 PIDs, both EOF, heavyReleased and all recorded processes absent. No worker or box was created.

Six focused tests and the released helper suite passed. Both locked full builds passed. The current tested shared production module is 2,203,481 bytes, SHA-256 eb1669a7d065664f8c3f0b3ae5649d9d222721b6c9903c125ef22aa5c21160ea. The final native fixture's tested Git HEAD is 5329f9982bb4b3d926f421feab90d67a5b37ce17; its exact foreign dirty hunks and owned new files are preserved as binary patches beside the pins. Production source was not edited, and unrelated NE/op5 work was neither committed nor adopted.

The final native-controls-2/ run uses the actual index, CSS, renderer, interpreter and GuestRPC with a small synthetic PE; no registered game launches. Current source/module/media/autoimports/TLB/fonts, actual resolved executables, native libraries, diagnostic compositions and all four historical comparison sources were pinned before launch. Historical sources are explicit red controls, never substituted for current qualification. All 119 HTTP responses were audited (116 successful), source/module/dependency pins had zero drift, and runtime errors were empty. The first successful fixture's narrower manifest omitted the explicit Git executable and three historical comparison-source pins; it is retained as limited evidence and does not replace the final complete run.

The final run retained 385 public operand observations. Actual controls passed:

Four personally reviewed final result-named fixture images are retained: host consent, fullscreen current output, native key display and native Mouse1 display. Other screenshots were pruned with hash/size metadata. No source, module/media archives, frame series or capture bundles were created. Actual native/watchdog/supervisor/outer and release-child waits, both EOF and absent recorded PIDs are retained in the runtime release receipt. Box state is NOT_CREATED; ID and expiry are null, stop is not applicable.

Remaining boundary

These controls do not reproduce the old original's failed transition. Successful synthetic controls are not proof that every original transition is correct. There is no warranted production change or conditional original authorization triggered by this work. A production/native fixture must first reproduce the specific refusal with its complete public operands and distinguish real occlusion, output routing, clipping and owner/input state. The prior original's screenshot-compositor UNKNOWN remains a separate missing cause. This bounded task stops without another unchanged original.